This action might not be possible to undo. Are you sure you want to continue?
Foundations and TrendsR_ in
Signal Processing
Vol. 1, Nos. 1–2 (2007) 1–194
_c 2007 L. R. Rabiner and R. W. Schafer
DOI: 10.1561/2000000001
Introduction to Digital Speech Processing
Lawrence R. Rabiner1 and Ronald W. Schafer2
1 Rutgers University and University of California, Santa Barbara,
USA,
rabiner@ece.ucsb.edu
2 HewlettPackard Laboratories, Palo Alto, CA, USA
Abstract
Since even before the time of Alexander Graham Bell’s
revolutionary
invention, engineers and scientists have studied the phenomenon
of speech communication with an eye on creating more efficient
and
effective systems of humantohuman and humantomachine
communication.
Starting in the 1960s, digital signal processing (DSP), assumed
a central role in speech studies, and today DSP is the key to
realizing
the fruits of the knowledge that has been gained through decades
of
research. Concomitant advances in integrated circuit technology
and
computer architecture have aligned to create a technological
environment
with virtually limitless opportunities for innovation in speech
communication applications. In this text, we highlight the central
role
of DSP techniques in modern speech communication research and
applications.
We present a comprehensive overview of digital speech processing
that ranges from the basic nature of the speech signal, through a
variety of methods of representing speech in digital form, to
applications
in voice communication and automatic synthesis and recognition
of speech. The breadth of this subject does not allow us to discuss
any
aspect of speech processing to great depth; hence our goal is to
provide
a useful introduction to the wide range of important concepts that
comprise the field of digital speech processing. A more
comprehensive
treatment will appear in the forthcoming book, Theory and
Application
of Digital Speech Processing [101].
1
Introduction
The fundamental purpose of speech is communication, i.e., the
transmission
of messages. According to Shannon’s information theory [116],
a message represented as a sequence of discrete symbols can be
quantified
by its information content in bits, and the rate of transmission of
information is measured in bits/second (bps). In speech production,
as
well as in many humanengineered electronic communication
systems,
the information to be transmitted is encoded in the form of a
continuously
varying (analog) waveform that can be transmitted, recorded,
manipulated, and ultimately decoded by a human listener. In the
case
of speech, the fundamental analog form of the message is an
acoustic
waveform, which we call the speech signal. Speech signals, as
illustrated
in Figure 1.1, can be converted to an electrical waveform by
a microphone, further manipulated by both analog and digital
signal
processing, and then converted back to acoustic form by a
loudspeaker,
a telephone handset or headphone, as desired. This form of speech
processing
is, of course, the basis for Bell’s telephone invention as well as
today’s multitude of devices for recording, transmitting, and
manipulating
speech and audio signals. Although Bell made his invention
without knowing the fundamentals of information theory, these
ideas
3
4 Introduction
0 0.02 0.04 0.06 0.08 0.1 0.12
0.48
0.36
0.24
0.12
0
time in seconds
SH UH
D W IY
CH
EY
S
Fig. 1.1 A speech waveform with phonetic labels for the text
message “Should we chase.”
have assumed great importance in the design of sophisticated
modern
communications systems. Therefore, even though our main focus
will
be mostly on the speech waveform and its representation in the
form of
parametric models, it is nevertheless useful to begin with a
discussion
of how information is encoded in the speech waveform.
1.1 The Speech Chain
Figure 1.2 shows the complete process of producing and
perceiving
speech from the formulation of a message in the brain of a talker,
to
the creation of the speech signal, and finally to the understanding
of
the message by a listener. In their classic introduction to speech
science,
Denes and Pinson aptly referred to this process as the “speech
chain” [29]. The process starts in the upper left as a message
represented
somehow in the brain of the speaker. The message information
can be thought of as having a number of different representations
during
the process of speech production (the upper path in Figure 1.2).
1.1 The Speech Chain 5
Message
Formulation
Message
Understanding
Language
Translation
Language
Code
NeuroMuscular
Controls
Vocal Tract
System
Transmission
Channel
acoustic
waveform
acoustic
waveform
Neural
Transduction
Basilar
Membrane
Motion
discrete input continuous input
Speech Production
Speech Perception
discrete output continuous output
text phonemes,
prosody articulatory motions
semantics
phonemes, words
sentences
feature
extraction
spectrum
analysis
50 bps 200 bps 2000 bps 64700 Kbps
information rate
excitation,
formants
Fig. 1.2 The Speech Chain: from message, to speech signal, to
understanding.
For example the message could be represented initially as English
text.
In order to “speak” the message, the talker implicitly converts the
text
into a symbolic representation of the sequence of sounds
corresponding
to the spoken version of the text. This step, called the language
code
generator in Figure 1.2, converts text symbols to phonetic symbols
(along with stress and durational information) that describe the
basic
sounds of a spoken version of the message and the manner (i.e., the
speed and emphasis) in which the sounds are intended to be
produced.
As an example, the segments of the waveform of Figure 1.1 are
labeled
with phonetic symbols using a computerkeyboardfriendly code
called
ARPAbet.1 Thus, the text “should we chase” is represented
phonetically
(in ARPAbet symbols) as [SH UH D — W IY — CH EY S]. (See
Chapter 2 for more discussion of phonetic transcription.) The third
step
in the speech production process is the conversion to “neuro
muscular
controls,” i.e., the set of control signals that direct the neuro
muscular
system to move the speech articulators, namely the tongue, lips,
teeth,
1 The International Phonetic Association (IPA) provides a set of
rules for phonetic transcription
using an equivalent set of specialized symbols. The ARPAbet code
does not
require special fonts and is thus more convenient for computer
applications.
6 Introduction
jaw and velum, in a manner that is consistent with the sounds of
the
desired spoken message and with the desired degree of emphasis.
The
end result of the neuromuscular controls step is a set of
articulatory
motions (continuous control) that cause the vocal tract articulators
to
move in a prescribed manner in order to create the desired sounds.
Finally the last step in the Speech Production process is the “vocal
tract system” that physically creates the necessary sound sources
and
the appropriate vocal tract shapes over time so as to create an
acoustic
waveform, such as the one shown in Figure 1.1, that encodes the
information
in the desired message into the speech signal.
To determine the rate of information flow during speech
production,
assume that there are about 32 symbols (letters) in the language
(in English there are 26 letters, but if we include simple
punctuation
we get a count closer to 32 = 25 symbols). Furthermore, the rate of
speaking for most people is about 10 symbols per second
(somewhat
on the high side, but still acceptable for a rough information rate
estimate).
Hence, assuming independent letters as a simple approximation,
we estimate the base information rate of the text message as about
50 bps (5 bits per symbol times 10 symbols per second). At the
second
stage of the process, where the text representation is converted into
phonemes and prosody (e.g., pitch and stress) markers, the
information
rate is estimated to increase by a factor of 4 to about 200 bps. For
example, the ARBAbet phonetic symbol set used to label the
speech
sounds in Figure 1.1 contains approximately 64 = 26 symbols, or
about
6 bits/phoneme (again a rough approximation assuming
independence
of phonemes). In Figure 1.1, there are 8 phonemes in
approximately
600 ms. This leads to an estimate of 8 × 6/0.6 = 80 bps. Additional
information required to describe prosodic features of the signal
(e.g.,
duration, pitch, loudness) could easily add 100 bps to the total
information
rate for a message encoded as a speech signal.
The information representations for the first two stages in the
speech
chain are discrete so we can readily estimate the rate of
information
flow with some simple assumptions. For the next stage in the
speech
production part of the speech chain, the representation becomes
continuous
(in the form of control signals for articulatory motion). If they
could be measured, we could estimate the spectral bandwidth of
these
1.1 The Speech Chain 7
control signals and appropriately sample and quantize these signals
to
obtain equivalent digital signals for which the data rate could be
estimated.
The articulators move relatively slowly compared to the time
variation of the resulting acoustic waveform. Estimates of
bandwidth
and required accuracy suggest that the total data rate of the
sampled
articulatory control signals is about 2000 bps [34]. Thus, the
original
text message is represented by a set of continuously varying
signals
whose digital representation requires a much higher data rate than
the information rate that we estimated for transmission of the
message
as a speech signal.2 Finally, as we will see later, the data rate
of the digitized speech waveform at the end of the speech
production
part of the speech chain can be anywhere from 64,000 to more than
700,000 bps. We arrive at such numbers by examining the
sampling
rate and quantization required to represent the speech signal with a
desired perceptual fidelity. For example, “telephone quality”
requires
that a bandwidth of 0–4 kHz be preserved, implying a sampling
rate of
8000 samples/s. Each sample can be quantized with 8 bits on a log
scale,
resulting in a bit rate of 64,000 bps. This representation is highly
intelligible
(i.e., humans can readily extract the message from it) but to most
listeners, it will sound different from the original speech signal
uttered
by the talker. On the other hand, the speech waveform can be
represented
with “CD quality” using a sampling rate of 44,100 samples/s
with 16 bit samples, or a data rate of 705,600 bps. In this case, the
reproduced acoustic signal will be virtually indistinguishable from
the
original speech signal.
As we move from text to speech waveform through the speech
chain,
the result is an encoding of the message that can be effectively
transmitted
by acoustic wave propagation and robustly decoded by the hearing
mechanism of a listener. The above analysis of data rates shows
that as we move from text to sampled speech waveform, the data
rate
can increase by a factor of 10,000. Part of this extra information
represents
characteristics of the talker such as emotional state, speech
mannerisms, accent, etc., but much of it is due to the inefficiency
2 Note that we introduce the term data rate for digital
representations to distinguish from
the inherent information content of the message represented by the
speech signal.
8 Introduction
of simply sampling and finely quantizing analog signals. Thus,
motivated
by an awareness of the low intrinsic information rate of speech,
a central theme of much of digital speech processing is to obtain a
digital representation with lower data rate than that of the sampled
waveform.
The complete speech chain consists of a speech production/
generation model, of the type discussed above, as well as a speech
perception/recognition model, as shown progressing to the left in
the
bottom half of Figure 1.2. The speech perception model shows the
series
of steps from capturing speech at the ear to understanding the
message
encoded in the speech signal. The first step is the effective
conversion
of the acoustic waveform to a spectral representation. This is
done within the inner ear by the basilar membrane, which acts as a
nonuniform spectrum analyzer by spatially separating the spectral
components of the incoming speech signal and thereby analyzing
them
by what amounts to a nonuniform filter bank. The next step in the
speech perception process is a neural transduction of the spectral
features
into a set of sound features (or distinctive features as they are
referred to in the area of linguistics) that can be decoded and
processed
by the brain. The next step in the process is a conversion of the
sound
features into the set of phonemes, words, and sentences associated
with
the incoming message by a language translation process in the
human
brain. Finally, the last step in the speech perception model is the
conversion
of the phonemes, words and sentences of the message into an
understanding of the meaning of the basic message in order to be
able
to respond to or take some appropriate action. Our fundamental
understanding
of the processes in most of the speech perception modules in
Figure 1.2 is rudimentary at best, but it is generally agreed that
some
physical correlate of each of the steps in the speech perception
model
occur within the human brain, and thus the entire model is useful
for
thinking about the processes that occur.
There is one additional process shown in the diagram of the
complete
speech chain in Figure 1.2 that we have not discussed — namely
the transmission channel between the speech generation and
speech
perception parts of the model. In its simplest embodiment, this
transmission
channel consists of just the acoustic wave connection between
1.2 Applications of Digital Speech Processing 9
a speaker and a listener who are in a common space. It is essential
to include this transmission channel in our model for the speech
chain since it includes real world noise and channel distortions that
make speech and message understanding more difficult in real
communication
environments. More interestingly for our purpose here —
it is in this domain that we find the applications of digital speech
processing.
1.2 Applications of Digital Speech Processing
The first step in most applications of digital speech processing is to
convert the acoustic waveform to a sequence of numbers. Most
modern
AtoD converters operate by sampling at a very high rate,
applying a
digital lowpass filter with cutoff set to preserve a prescribed
bandwidth,
and then reducing the sampling rate to the desired sampling rate,
which
can be as low as twice the cutoff frequency of the sharpcutoff
digital
filter. This discretetime representation is the starting point for
most
applications. From this point, other representations are obtained by
digital processing. For the most part, these alternative
representations
are based on incorporating knowledge about the workings of the
speech
chain as depicted in Figure 1.2. As we will see, it is possible to
incorporate
aspects of both the speech production and speech perception
process into the digital representation and processing. It is not an
oversimplification
to assert that digital speech processing is grounded in a
set of techniques that have the goal of pushing the data rate of the
speech representation to the left along either the upper or lower
path
in Figure 1.2.
The remainder of this chapter is devoted to a brief summary of the
applications of digital speech processing, i.e., the systems that
people
interact with daily. Our discussion will confirm the importance of
the
digital representation in all application areas.
1.2.1 Speech Coding
Perhaps the most widespread applications of digital speech
processing
technology occur in the areas of digital transmission and storage
10 Introduction
AtoD
Converter
Analysis/
Encoding
Channel or
Medium
Synthesis/
Decoding
DtoA
Converter
speech
signal samples data
samples data
xc (t) x[n] y[n]
xc (t) x[n] y[n]
decoded
signal
Fig. 1.3 Speech coding block diagram — encoder and decoder.
of speech signals. In these areas the centrality of the digital
representation
is obvious, since the goal is to compress the digital waveform
representation of speech into a lower bitrate representation. It
is common to refer to this activity as “speech coding” or “speech
compression.”
Figure 1.3 shows a block diagram of a generic speech encoding/
decoding (or compression) system. In the upper part of the figure,
the AtoD converter converts the analog speech signal xc(t) to a
sampled
waveform representation x[n]. The digital signal x[n] is analyzed
and coded by digital computation algorithms to produce a new
digital
signal y[n] that can be transmitted over a digital communication
channel
or stored in a digital storage medium as ˆy[n]. As we will see, there
are a myriad of ways to do the encoding so as to reduce the data
rate
over that of the sampled and quantized speech waveform x[n].
Because
the digital representation at this point is often not directly related to
the sampled speech waveform, y[n] and ˆy[n] are appropriately
referred
to as data signals that represent the speech signal. The lower path
in
Figure 1.3 shows the decoder associated with the speech coder.
The
received data signal ˆy[n] is decoded using the inverse of the
analysis
processing, giving the sequence of samples ˆx[n] which is then
converted
(using a DtoA Converter) back to an analog signal ˆxc(t) for
human
listening. The decoder is often called a synthesizer because it must
reconstitute the speech waveform from data that may bear no direct
relationship to the waveform.
1.2 Applications of Digital Speech Processing 11
With carefully designed error protection coding of the digital
representation, the transmitted (y[n]) and received (ˆy[n]) data can
be
essentially identical. This is the quintessential feature of digital
coding.
In theory, perfect transmission of the coded digital representation
is
possible even under very noisy channel conditions, and in the case
of
digital storage, it is possible to store a perfect copy of the digital
representation
in perpetuity if sufficient care is taken to update the storage
medium as storage technology advances. This means that the
speech
signal can be reconstructed to within the accuracy of the original
coding
for as long as the digital representation is retained. In either case,
the goal of the speech coder is to start with samples of the speech
signal
and reduce (compress) the data rate required to represent the
speech
signal while maintaining a desired perceptual fidelity. The
compressed
representation can be more efficiently transmitted or stored, or the
bits
saved can be devoted to error protection.
Speech coders enable a broad range of applications including
narrowband
and broadband wired telephony, cellular communications,
voice over internet protocol (VoIP) (which utilizes the internet as
a realtime communications medium), secure voice for privacy and
encryption (for national security applications), extremely
narrowband
communications channels (such as battlefield applications using
high
frequency (HF) radio), and for storage of speech for telephone
answering
machines, interactive voice response (IVR) systems, and
prerecorded
messages. Speech coders often utilize many aspects of both
the speech production and speech perception processes, and hence
may
not be useful for more general audio signals such as music. Coders
that
are based on incorporating only aspects of sound perception
generally
do not achieve as much compression as those based on speech
production,
but they are more general and can be used for all types of audio
signals. These coders are widely deployed in MP3 and AAC
players and
for audio in digital television systems [120].
1.2.2 TexttoSpeech Synthesis
For many years, scientists and engineers have studied the speech
production
process with the goal of building a system that can start with
12 Introduction
Linguistic
Rules
Synthesis
Algorithm
DtoA
Converter
text speech
Fig. 1.4 Texttospeech synthesis system block diagram.
text and produce speech automatically. In a sense, a texttospeech
synthesizer such as depicted in Figure 1.4 is a digital simulation of
the
entire upper part of the speech chain diagram. The input to the
system
is ordinary text such as an email message or an article from a
newspaper
or magazine. The first block in the texttospeech synthesis system,
labeled linguistic rules, has the job of converting the printed text
input
into a set of sounds that the machine must synthesize. The
conversion
from text to sounds involves a set of linguistic rules that must
determine
the appropriate set of sounds (perhaps including things like
emphasis,
pauses, rates of speaking, etc.) so that the resulting synthetic
speech
will express the words and intent of the text message in what
passes
for a natural voice that can be decoded accurately by human
speech
perception. This is more difficult than simply looking up the words
in
a pronouncing dictionary because the linguistic rules must
determine
how to pronounce acronyms, how to pronounce ambiguous words
like
read, bass, object, how to pronounce abbreviations like St. (street
or
Saint), Dr. (Doctor or drive), and how to properly pronounce
proper
names, specialized terms, etc. Once the proper pronunciation of the
text has been determined, the role of the synthesis algorithm is to
create
the appropriate sound sequence to represent the text message in
the form of speech. In essence, the synthesis algorithm must
simulate
the action of the vocal tract system in creating the sounds of
speech.
There are many procedures for assembling the speech sounds and
compiling
them into a proper sentence, but the most promising one today is
called “unit selection and concatenation.” In this method, the
computer
stores multiple versions of each of the basic units of speech
(phones, half
phones, syllables, etc.), and then decides which sequence of speech
units
sounds best for the particular text message that is being produced.
The
basic digital representation is not generally the sampled speech
wave.
Instead, some sort of compressed representation is normally used
to
1.2 Applications of Digital Speech Processing 13
save memory and, more importantly, to allow convenient
manipulation
of durations and blending of adjacent sounds. Thus, the speech
synthesis
algorithm would include an appropriate decoder, as discussed in
Section 1.2.1, whose output is converted to an analog
representation
via the DtoA converter.
Texttospeech synthesis systems are an essential component of
modern human–machine communications systems and are used to
do
things like read email messages over a telephone, provide voice
output
from GPS systems in automobiles, provide the voices for talking
agents for completion of transactions over the internet, handle call
center
help desks and customer care applications, serve as the voice for
providing information from handheld devices such as foreign
language
phrasebooks, dictionaries, crossword puzzle helpers, and as the
voice of
announcement machines that provide information such as stock
quotes,
airline schedules, updates on arrivals and departures of flights, etc.
Another important application is in reading machines for the blind,
where an optical character recognition system provides the text
input
to a speech synthesis system.
1.2.3 Speech Recognition and Other Pattern
Matching Problems
Another large class of digital speech processing applications is
concerned
with the automatic extraction of information from the speech
signal. Most such systems involve some sort of pattern matching.
Figure 1.5 shows a block diagram of a generic approach to pattern
matching problems in speech processing. Such problems include
the
following: speech recognition, where the object is to extract the
message
from the speech signal; speaker recognition, where the goal is
to identify who is speaking; speaker verification, where the goal is
to verify a speaker’s claimed identity from analysis of their speech
AtoD
Converter
Feature
Analysis
Pattern
Matching
speech symbols
Fig. 1.5 Block diagram of general pattern matching system for
speech signals.
14 Introduction
signal; word spotting, which involves monitoring a speech signal
for
the occurrence of specified words or phrases; and automatic
indexing
of speech recordings based on recognition (or spotting) of spoken
keywords.
The first block in the pattern matching system converts the analog
speech waveform to digital form using an AtoD converter. The
feature analysis module converts the sampled speech signal to a set
of feature vectors. Often, the same analysis techniques that are
used
in speech coding are also used to derive the feature vectors. The
final
block in the system, namely the pattern matching block,
dynamically
time aligns the set of feature vectors representing the speech signal
with
a concatenated set of stored patterns, and chooses the identity
associated
with the pattern which is the closest match to the timealigned set
of feature vectors of the speech signal. The symbolic output
consists
of a set of recognized words, in the case of speech recognition, or
the
identity of the best matching talker, in the case of speaker
recognition,
or a decision as to whether to accept or reject the identity claim of
a
speaker in the case of speaker verification.
Although the block diagram of Figure 1.5 represents a wide range
of speech pattern matching problems, the biggest use has been in
the
area of recognition and understanding of speech in support of
human–
machine communication by voice. The major areas where such a
system
finds applications include command and control of computer
software,
voice dictation to create letters, memos, and other documents,
natural
language voice dialogues with machines to enable help desks and
call centers, and for agent services such as calendar entry and
update,
address list modification and entry, etc.
Pattern recognition applications often occur in conjunction with
other digital speech processing applications. For example, one of
the
preeminent uses of speech technology is in portable
communication
devices. Speech coding at bit rates on the order of 8 Kbps enables
normal
voice conversations in cell phones. Spoken name speech
recognition
in cellphones enables voice dialing capability that can
automatically
dial the number associated with the recognized name. Names from
directories with upwards of several hundred names can readily be
recognized
and dialed using simple speech recognition technology.
1.2 Applications of Digital Speech Processing 15
Another major speech application that has long been a dream
of speech researchers is automatic language translation. The goal
of
language translation systems is to convert spoken words in one
language
to spoken words in another language so as to facilitate natural
language voice dialogues between people speaking different
languages.
Language translation technology requires speech synthesis systems
that
work in both languages, along with speech recognition (and
generally
natural language understanding) that also works for both
languages;
hence it is a very difficult task and one for which only limited
progress
has been made. When such systems exist, it will be possible for
people
speaking different languages to communicate at data rates on the
order
of that of printed text reading!
1.2.4 Other Speech Applications
The range of speech communication applications is illustrated in
Figure 1.6. As seen in this figure, the techniques of digital speech
processing are a key ingredient of a wide range of applications that
include the three areas of transmission/storage, speech synthesis,
and
speech recognition as well as many others such as speaker
identification,
speech signal quality enhancement, and aids for the hearing or
visuallyimpaired.
The block diagram in Figure 1.7 represents any system where time
signals such as speech are processed by the techniques of DSP.
This
figure simply depicts the notion that once the speech signal is
sampled,
it can be manipulated in virtually limitless ways by DSP
techniques.
Here again, manipulations and modifications of the speech signal
are
Digital
Transmission
& Storage
Digital Speech Processing
Techniques
Speech
Synthesis
Speech
Recognition
Speaker
Verification/
Identification
Enhancement
of Speech
Quality
Aids for the
Handicapped
Fig. 1.6 Range of speech communication applications.
16 Introduction
AtoD
Converter
Computer
Algorithm
DtoA
Converter
speech speech
Fig. 1.7 General block diagram for application of digital signal
processing to speech signals.
usually achieved by transforming the speech signal into an
alternative
representation (that is motivated by our understanding of speech
production
and speech perception), operating on that representation by
further digital computation, and then transforming back to the
waveform
domain, using a DtoA converter.
One important application area is speech enhancement, where the
goal is to remove or suppress noise or echo or reverberation picked
up by
a microphone along with the desired speech signal. In humanto
human
communication, the goal of speech enhancement systems is to
make the
speech more intelligible and more natural; however, in reality the
best
that has been achieved so far is less perceptually annoying speech
that
essentially maintains, but does not improve, the intelligibility of
the
noisy speech. Success has been achieved, however, in making
distorted
speech signals more useful for further processing as part of a
speech
coder, synthesizer, or recognizer. An excellent reference in this
area is
the recent textbook by Loizou [72].
Other examples of manipulation of the speech signal include
timescale modification to align voices with video segments, to
modify
voice qualities, and to speedup or slowdown prerecorded speech
(e.g., for talking books, rapid review of voice mail messages, or
careful
scrutinizing of spoken material).
1.3 Our Goal for this Text
We have discussed the speech signal and how it encodes
information
for human communication. We have given a brief overview of the
way
in which digital speech processing is being applied today, and we
have
hinted at some of the possibilities that exist for the future. These
and
many more examples all rely on the basic principles of digital
speech
processing, which we will discuss in the remainder of this text. We
make no pretense of exhaustive coverage. The subject is too broad
and
1.3 Our Goal for this Text 17
too deep. Our goal is only to provide an uptodate introduction to
this
fascinating field. We will not be able to go into great depth, and we
will
not be able to cover all the possible applications of digital speech
processing
techniques. Instead our focus is on the fundamentals of digital
speech processing and their application to coding, synthesis, and
recognition.
This means that some of the latest algorithmic innovations and
applications will not be discussed — not because they are not
interesting,
but simply because there are so many fundamental triedandtrue
techniques that remain at the core of digital speech processing. We
hope that this text will stimulate readers to investigate the subject
in
greater depth using the extensive set of references provided.
2
The Speech Signal
As the discussion in Chapter 1 shows, the goal in many
applications of
digital speech processing techniques is to move the digital
representation
of the speech signal from the waveform samples back up the
speech
chain toward the message. To gain a better idea of what this
means,
this chapter provides a brief overview of the phonetic
representation of
speech and an introduction to models for the production of the
speech
signal.
2.1 Phonetic Representation of Speech
Speech can be represented phonetically by a finite set of symbols
called
the phonemes of the language, the number of which depends upon
the
language and the refinement of the analysis. For most languages
the
number of phonemes is between 32 and 64. A condensed inventory
of
the sounds of speech in the English language is given in Table 2.1,
where the phonemes are denoted by a set of ASCII symbols called
the
ARPAbet. Table 2.1 also includes some simple examples of
ARPAbet
transcriptions of words containing each of the phonemes of
English.
Additional phonemes can be added to Table 2.1 to account for
allophonic
variations and events such as glottal stops and pauses.
18
2.1 Phonetic Representation of Speech 19
Table 2.1 Condensed list of ARPAbet phonetic symbols for North
American English.
Class ARPAbet Example Transcription
Vowels and IY beet [B IY T]
diphthongs IH bit [B IH T]
EY bait [B EY T]
EH bet [B EH T]
AE bat [B AE T]
AA bob [B AA B]
AO born [B AO R N]
UH book [B UH K]
OW boat [B OW T]
UW boot [B UW T]
AH but [B AH T ]
ER bird [B ER D]
AY buy [B AY]
AW down [D AW N]
OY boy [B OY]
Glides Y you [Y UH]
R rent [R EH N T]
Liquids W wit [W IH T]
L let [L EH T]
Nasals M met [M EH T]
N net [N EH T]
NG sing [S IH NG]
Stops P pat [P AE T]
B bet [B EH T]
T ten [T EH N]
D debt [D EH T]
K kit [K IH T]
G get [G EH T]
Fricatives HH hat [HH AE T]
F f at [F AE T]
V vat [V AE T]
TH thing [TH IH NG]
DH that [DH AE T]
S sat [S AE T]
Z zoo [Z UW]
SH shut [SH AH T]
ZH az ure [AE ZH ER]
Affricates CH chase [CH EY S]
JH j udge [JH AH JH ]
aThis set of 39 phonemes is used in the CMU Pronouncing
Dictionary available online at
http://www.speech.cs.cmu.edu/cgibin/cmudict.
Figure 1.1 on p. 4 shows how the sounds corresponding to the text
“should we chase” are encoded into a speech waveform. We see
that, for
the most part, phonemes have a distinctive appearance in the
speech
waveform. Thus sounds like /SH/ and /S/ look like (spectrally
shaped)
20 The Speech Signal
random noise, while the vowel sounds /UH/, /IY/, and /EY/ are
highly
structured and quasiperiodic. These differences result from the
distinctively
different ways that these sounds are produced.
2.2 Models for Speech Production
A schematic longitudinal crosssectional drawing of the human
vocal
tract mechanism is given in Figure 2.1 [35]. This diagram
highlights
the essential physical features of human anatomy that enter into the
final stages of the speech production process. It shows the vocal
tract
as a tube of nonuniform crosssectional area that is bounded at one
end
by the vocal cords and at the other by the mouth opening. This tube
serves as an acoustic transmission system for sounds generated
inside
the vocal tract. For creating nasal sounds like /M/, /N/, or /NG/,
a sidebranch tube, called the nasal tract, is connected to the main
acoustic branch by the trapdoor action of the velum. This branch
path
radiates sound at the nostrils. The shape (variation of crosssection
along the axis) of the vocal tract varies with time due to motions of
the lips, jaw, tongue, and velum. Although the actual human vocal
tract is not laid out along a straight line as in Figure 2.1, this type
of
model is a reasonable approximation for wavelengths of the sounds
in
speech.
The sounds of speech are generated in the system of Figure 2.1 in
several ways. Voiced sounds (vowels, liquids, glides, nasals in
Table 2.1)
Fig. 2.1 Schematic model of the vocal tract system. (After
Flanagan et al. [35].)
2.2 Models for Speech Production 21
are produced when the vocal tract tube is excited by pulses of air
pressure
resulting from quasiperiodic opening and closing of the glottal
orifice (opening between the vocal cords). Examples in Figure 1.1
are
the vowels /UH/, /IY/, and /EY/, and the liquid consonant /W/.
Unvoiced sounds are produced by creating a constriction
somewhere in
the vocal tract tube and forcing air through that constriction,
thereby
creating turbulent air flow, which acts as a random noise excitation
of
the vocal tract tube. Examples are the unvoiced fricative sounds
such as
/SH/ and /S/. A third sound production mechanism is when the
vocal
tract is partially closed off causing turbulent flow due to the
constriction,
at the same time allowing quasiperiodic flow due to vocal cord
vibrations. Sounds produced in this manner include the voiced
fricatives
/V/, /DH/, /Z/, and /ZH/. Finally, plosive sounds such as /P/,
/T/, and /K/ and affricates such as /CH/ are formed by momentarily
closing off air flow, allowing pressure to build up behind the
closure, and
then abruptly releasing the pressure. All these excitation sources
create
a wideband excitation signal to the vocal tract tube, which acts as
an acoustic transmission line with certain vocal tract shape
dependent
resonances that tend to emphasize some frequencies of the
excitation
relative to others.
As discussed in Chapter 1 and illustrated by the waveform in
Figure 1.1, the general character of the speech signal varies at the
phoneme rate, which is on the order of 10 phonemes per second,
while
the detailed time variations of the speech waveform are at a much
higher
rate. That is, the changes in vocal tract configuration occur
relatively
slowly compared to the detailed time variation of the speech
signal. The
sounds created in the vocal tract are shaped in the frequency
domain
by the frequency response of the vocal tract. The resonance
frequencies
resulting from a particular configuration of the articulators are
instrumental
in forming the sound corresponding to a given phoneme. These
resonance frequencies are called the formant frequencies of the
sound
[32, 34]. In summary, the fine structure of the time waveform is
created
by the sound sources in the vocal tract, and the resonances of the
vocal
tract tube shape these sound sources into the phonemes.
The system of Figure 2.1 can be described by acoustic theory,
and numerical techniques can be used to create a complete physical
22 The Speech Signal
Linear
System
Excitation
Generator excitation signal speech signal
Vocal Tract
Parameters
Excitation
Parameters
Fig. 2.2 Source/system model for a speech signal.
simulation of sound generation and transmission in the vocal tract
[36, 93], but, for the most part, it is sufficient to model the
production
of a sampled speech signal by a discretetime system model such
as the one depicted in Figure 2.2. The discretetime timevarying
linear
system on the right in Figure 2.2 simulates the frequency shaping
of
the vocal tract tube. The excitation generator on the left simulates
the
different modes of sound generation in the vocal tract. Samples of
a
speech signal are assumed to be the output of the timevarying
linear
system.
In general such a model is called a source/system model of speech
production. The shorttime frequency response of the linear system
simulates the frequency shaping of the vocal tract system, and
since the
vocal tract changes shape relatively slowly, it is reasonable to
assume
that the linear system response does not vary over time intervals on
the
order of 10 ms or so. Thus, it is common to characterize the
discretetime
linear system by a system function of the form:
H(z) =
M_
k=0
bkz
−k
1 −
_N
k=1
akz
−k
=
b0
M_
k=1
(1 − dkz
−1)
N_
k=1
(1 − ckz
−1)
, (2.1)
where the filter coefficients ak and bk (labeled as vocal tract
parameters
in Figure 2.2) change at a rate on the order of 50–100 times/s.
Some
of the poles (ck) of the system function lie close to the unit circle
and create resonances to model the formant frequencies. In detailed
modeling of speech production [32, 34, 64], it is sometimes useful
to
employ zeros (dk) of the system function to model nasal and
fricative
2.2 Models for Speech Production 23
sounds. However, as we discuss further in Chapter 4, many
applications
of the source/system model only include poles in the model
because this
simplifies the analysis required to estimate the parameters of the
model
from the speech signal.
The box labeled excitation generator in Figure 2.2 creates an
appropriate
excitation for the type of sound being produced. For voiced
speech the excitation to the linear system is a quasiperiodic
sequence
of discrete (glottal) pulses that look very much like those shown in
the
righthand half of the excitation signal waveform in Figure 2.2. The
fundamental
frequency of the glottal excitation determines the perceived
pitch of the voice. The individual finiteduration glottal pulses
have a
lowpass spectrum that depends on a number of factors [105].
Therefore,
the periodic sequence of smooth glottal pulses has a harmonic line
spectrum with components that decrease in amplitude with
increasing
frequency. Often it is convenient to absorb the glottal pulse
spectrum
contribution into the vocal tract system model of (2.1). This can be
achieved by a small increase in the order of the denominator over
what
would be needed to represent the formant resonances. For
unvoiced
speech, the linear system is excited by a random number generator
that produces a discretetime noise signal with flat spectrum as
shown
in the lefthand half of the excitation signal. The excitation in
Figure
2.2 switches from unvoiced to voiced leading to the speech signal
output as shown in the figure. In either case, the linear system
imposes
its frequency response on the spectrum to create the speech sounds.
This model of speech as the output of a slowly timevarying digital
filter with an excitation that captures the nature of the
voiced/unvoiced
distinction in speech production is the basis for thinking about the
speech signal, and a wide variety of digital representations of the
speech
signal are based on it. That is, the speech signal is represented by
the
parameters of the model instead of the sampled waveform. By
assuming
that the properties of the speech signal (and the model) are
constant
over short time intervals, it is possible to
compute/measure/estimate
the parameters of the model by analyzing short blocks of samples
of
the speech signal. It is through such models and analysis
techniques
that we are able to build properties of the speech production
process
into digital representations of the speech signal.
24 The Speech Signal
2.3 More Refined Models
Source/system models as shown in Figure 2.2 with the system
characterized
by a timesequence of timeinvariant systems are quite sufficient
for most applications in speech processing, and we shall rely on
such models throughout this text. However, such models are based
on
many approximations including the assumption that the source and
the
system do not interact, the assumption of linearity, and the
assumption
that the distributed continuoustime vocal tract transmission
system
can be modeled by a discrete linear timeinvariant system. Fluid
mechanics and acoustic wave propagation theory are fundamental
physical
principles that must be applied for detailed modeling of speech
production. Since the early work of Flanagan and Ishizaka [34, 36,
51]
much work has been devoted to creating detailed simulations of
glottal
flow, the interaction of the glottal source and the vocal tract in
speech
production, and the nonlinearities that enter into sound generation
and
transmission in the vocal tract. Stevens [121] and Quatieri [94]
provide
useful discussions of these effects. For many years, researchers
have
sought to measure the physical dimensions of the human vocal
tract
during speech production. This information is essential for detailed
simulations
based on acoustic theory. Early efforts to measure vocal tract
area functions involved hand tracing on Xray pictures [32].
Recent
advances in MRI imaging and computer image analysis have
provided
significant advances in this area of speech science [17].
3
Hearing and Auditory Perception
In Chapter 2, we introduced the speech production process and
showed
how we could model speech production using discretetime
systems.
In this chapter we turn to the perception side of the speech chain to
discuss properties of human sound perception that can be
employed to
create digital representations of the speech signal that are
perceptually
robust.
3.1 The Human Ear
Figure 3.1 shows a schematic view of the human ear showing the
three
distinct sound processing sections, namely: the outer ear
consisting of
the pinna, which gathers sound and conducts it through the
external
canal to the middle ear; the middle ear beginning at the tympanic
membrane, or eardrum, and including three small bones, the
malleus
(also called the hammer), the incus (also called the anvil) and the
stapes
(also called the stirrup), which perform a transduction from
acoustic
waves to mechanical pressure waves; and finally, the inner ear,
which
consists of the cochlea and the set of neural connections to the
auditory
nerve, which conducts the neural signals to the brain.
25
26 Hearing and Auditory Perception
Fig. 3.1 Schematic view of the human ear (inner and middle
structures enlarged). (After
Flanagan [34].)
Figure 3.2 [107] depicts a block diagram abstraction of the
auditory
system. The acoustic wave is transmitted from the outer ear to the
inner ear where the ear drum and bone structures convert the sound
wave to mechanical vibrations which ultimately are transferred to
the
basilar membrane inside the cochlea. The basilar membrane
vibrates in
a frequencyselective manner along its extent and thereby performs
a
rough (nonuniform) spectral analysis of the sound. Distributed
along
Fig. 3.2 Schematic model of the auditory mechanism. (After Sachs
et al. [107].)
3.2 Perception of Loudness 27
the basilar membrane are a set of inner hair cells that serve to
convert
motion along the basilar membrane to neural activity. This
produces
an auditory nerve representation in both time and frequency. The
processing
at higher levels in the brain, shown in Figure 3.2 as a sequence
of central processing with multiple representations followed by
some
type of pattern recognition, is not well understood and we can only
postulate the mechanisms used by the human brain to perceive
sound
or speech. Even so, a wealth of knowledge about how sounds are
perceived
has been discovered by careful experiments that use tones and
noise signals to stimulate the auditory system of human observers
in
very specific and controlled ways. These experiments have yielded
much
valuable knowledge about the sensitivity of the human auditory
system
to acoustic properties such as intensity and frequency.
3.2 Perception of Loudness
A key factor in the perception of speech and other sounds is
loudness.
Loudness is a perceptual quality that is related to the physical
property
of sound pressure level. Loudness is quantified by relating the
actual
sound pressure level of a pure tone (in dB relative to a standard
reference
level) to the perceived loudness of the same tone (in a unit called
phons) over the range of human hearing (20 Hz–20 kHz). This
relationship
is shown in Figure 3.3 [37, 103]. These loudness curves show that
the perception of loudness is frequencydependent. Specifically,
the dotted
curve at the bottom of the figure labeled “threshold of audibility”
shows the sound pressure level that is required for a sound of a
given
frequency to be just audible (by a person with normal hearing). It
can
be seen that low frequencies must be significantly more intense
than
frequencies in the midrange in order that they be perceived at all.
The
solid curves are equalloudnesslevel contours measured by
comparing
sounds at various frequencies with a pure tone of frequency 1000
Hz and
known sound pressure level. For example, the point at frequency
100 Hz
on the curve labeled 50 (phons) is obtained by adjusting the power
of
the 100 Hz tone until it sounds as loud as a 1000 Hz tone having a
sound
pressure level of 50 dB. Careful measurements of this kind show
that a
100 Hz tone must have a sound pressure level of about 60 dB in
order
28 Hearing and Auditory Perception
Fig. 3.3 Loudness level for human hearing. (After Fletcher and
Munson [37].)
to be perceived to be equal in loudness to the 1000 Hz tone of
sound
pressure level 50 dB. By convention, both the 50 dB 1000 Hz tone
and
the 60 dB 100 Hz tone are said to have a loudness level of 50
phons
(pronounced as /F OW N Z/).
The equalloudnesslevel curves show that the auditory system is
most sensitive for frequencies ranging from about 100 Hz up to
about
6 kHz with the greatest sensitivity at around 3 to 4 kHz. This is
almost
precisely the range of frequencies occupied by most of the sounds
of
speech.
3.3 Critical Bands
The nonuniform frequency analysis performed by the basilar
membrane
can be thought of as equivalent to that of a set of bandpass filters
whose frequency responses become increasingly broad with
increasing
frequency. An idealized version of such a filter bank is depicted
schematically in Figure 3.4. In reality, the bandpass filters are not
ideal
3.4 Pitch Perception 29
amplitude
Fig. 3.4 Schematic representation of bandpass filters according to
the critical band theory
of hearing.
as shown in Figure 3.4, but their frequency responses overlap
significantly
since points on the basilar membrane cannot vibrate independently
of each other. Even so, the concept of bandpass filter analysis
in the cochlea is well established, and the critical bandwidths have
been defined and measured using a variety of methods, showing
that
the effective bandwidths are constant at about 100 Hz for center
frequencies
below 500 Hz, and with a relative bandwidth of about 20%
of the center frequency above 500 Hz. An equation that fits
empirical
measurements over the auditory range is
Δfc =25 + 75[1 + 1.4(fc/1000)2]0.69, (3.1)
where Δfc is the critical bandwidth associated with center
frequency
fc [134]. Approximately 25 critical band filters span the range from
0 to 20 kHz. The concept of critical bands is very important in
understanding
such phenomena as loudness perception, pitch perception, and
masking, and it therefore provides motivation for digital
representations
of the speech signal that are based on a frequency decomposition.
3.4 Pitch Perception
Most musical sounds as well as voiced speech sounds have a
periodic
structure when viewed over short time intervals, and such sounds
are
perceived by the auditory system as having a quality known as
pitch.
Like loudness, pitch is a subjective attribute of sound that is related
to
the fundamental frequency of the sound, which is a physical
attribute of
the acoustic waveform [122]. The relationship between pitch
(measured
30 Hearing and Auditory Perception
Fig. 3.5 Relation between subjective pitch and frequency of a pure
tone.
on a nonlinear frequency scale called the melscale) and frequency
of a
pure tone is approximated by the equation [122]:
Pitch in mels = 1127loge(1 + f/700), (3.2)
which is plotted in Figure 3.5. This expression is calibrated so that
a
frequency of 1000 Hz corresponds to a pitch of 1000 mels. This
empirical
scale describes the results of experiments where subjects were
asked to
adjust the pitch of a measurement tone to half the pitch of a
reference
tone. To calibrate the scale, a tone of frequency 1000 Hz is given a
pitch of 1000 mels. Below 1000 Hz, the relationship between pitch
and
frequency is nearly proportional. For higher frequencies, however,
the
relationship is nonlinear. For example, (3.2) shows that a
frequency of
f = 5000 Hz corresponds to a pitch of 2364 mels.
The psychophysical phenomenon of pitch, as quantified by the
melscale,
can be related to the concept of critical bands [134]. It turns out
that more or less independently of the center frequency of the
band, one
critical bandwidth corresponds to about 100 mels on the pitch
scale.
This is shown in Figure 3.5, where a critical band of width Δfc =
160 Hz
centered on fc = 1000 Hz maps into a band of width 106 mels and
a
critical band of width 100 Hz centered on 350 Hz maps into a band
of
width 107 mels. Thus, what we know about pitch perception
reinforces
the notion that the auditory system performs a frequency analysis
that
can be simulated with a bank of bandpass filters whose bandwidths
increase as center frequency increases.
3.5 Auditory Masking 31
Voiced speech is quasiperiodic, but contains many frequencies.
Nevertheless,
many of the results obtained with pure tones are relevant to
the perception of voice pitch as well. Often the term pitch period is
used
for the fundamental period of the voiced speech signal even though
its
usage in this way is somewhat imprecise.
3.5 Auditory Masking
The phenomenon of critical band auditory analysis can be
explained
intuitively in terms of vibrations of the basilar membrane. A
related
phenomenon, called masking, is also attributable to the mechanical
vibrations of the basilar membrane. Masking occurs when one
sound
makes a second superimposed sound inaudible. Loud tones causing
strong vibrations at a point on the basilar membrane can swamp
out
vibrations that occur nearby. Pure tones can mask other pure tones,
and noise can mask pure tones as well. A detailed discussion of
masking
can be found in [134].
Figure 3.6 illustrates masking of tones by tones. The notion that a
sound becomes inaudible can be quantified with respect to the
threshold
of audibility. As shown in Figure 3.6, an intense tone (called the
masker) tends to raise the threshold of audibility around its
location on
the frequency axis as shown by the solid line. All spectral
components
whose level is below this raised threshold are masked and therefore
do
Masker
Shifted
threshold
Threshold
of hearing
Masked
signals
(dashed)
Unmasked
signal
log frequency
amplitude
Fig. 3.6 Illustration of effects of masking.
32 Hearing and Auditory Perception
not need to be reproduced in a speech (or audio) processing system
because they would not be heard. Similarly, any spectral
component
whose level is above the raised threshold is not masked, and
therefore
will be heard. It has been shown that the masking effect is greater
for
frequencies above the masking frequency than below. This is
shown in
Figure 3.6 where the falloff of the shifted threshold is less abrupt
above
than below the masker.
Masking is widely employed in digital representations of speech
(and
audio) signals by “hiding” errors in the representation in areas
where
the threshold of hearing is elevated by strong frequency
components in
the signal [120]. In this way, it is possible to achieve lower data
rate
representations while maintaining a high degree of perceptual
fidelity.
3.6 Complete Model of Auditory Processing
In Chapter 2, we described elements of a generative model of
speech
production which could, in theory, completely describe and model
the
ways in which speech is produced by humans. In this chapter, we
described elements of a model of speech perception. However, the
problem that arises is that our detailed knowledge of how speech is
perceived and understood, beyond the basilar membrane
processing
of the inner ear, is rudimentary at best, and thus we rely on
psychophysical
experimentation to understand the role of loudness, critical
bands, pitch perception, and auditory masking in speech perception
in
humans. Although some excellent auditory models have been
proposed
[41, 48, 73, 115] and used in a range of speech processing systems,
all
such models are incomplete representations of our knowledge
about
how speech is understood.
4
ShortTime Analysis of Speech
In Figure 2.2 of Chapter 2, we presented a model for speech
production
in which an excitation source provides the basic temporal fine
structure while a slowly varying filter provides spectral shaping
(often
referred to as the spectrum envelope) to create the various sounds
of
speech. In Figure 2.2, the source/system separation was presented
at
an abstract level, and few details of the excitation or the linear
system
were given. Both the excitation and the linear system were defined
implicitly by the assertion that the sampled speech signal was the
output
of the overall system. Clearly, this is not sufficient to uniquely
specify either the excitation or the system. Since our goal is to
extract
parameters of the model by analysis of the speech signal, it is
common
to assume structures (or representations) for both the excitation
generator and the linear system. One such model uses a more
detailed
representation of the excitation in terms of separate source
generators
for voiced and unvoiced speech as shown in Figure 4.1. In this
model
the unvoiced excitation is assumed to be a random noise sequence,
and the voiced excitation is assumed to be a periodic impulse train
with impulses spaced by the pitch period (P0) rounded to the
nearest
33
34 ShortTime Analysis of Speech
Fig. 4.1 Voiced/unvoiced/system model for a speech signal.
sample.1 The pulses needed to model the glottal flow waveform
during
voiced speech are assumed to be combined (by convolution) with
the impulse response of the linear system, which is assumed to be
slowlytimevarying (changing every 50–100 ms or so). By this we
mean
that over the timescale of phonemes, the impulse response,
frequency
response, and system function of the system remains relatively
constant.
For example over time intervals of tens of milliseconds, the system
can be described by the convolution expression
sˆn[n] =
∞_
m=0
hˆn[m]eˆn[n − m], (4.1)
where the subscript ˆn denotes the time index pointing to the block
of
samples of the entire speech signal s[n] wherein the impulse
response
hˆn[m] applies.We use n for the time index within that interval, and
m is
the index of summation in the convolution sum. In this model, the
gain
Gˆn is absorbed into hˆn[m] for convenience. As discussed in
Chapter 2,
the most general linear timeinvariant system would be
characterized
by a rational system function as given in (2.1). However, to
simplify
analysis, it is often assumed that the system is an allpole system
with
1 As mentioned in Chapter 3, the period of voiced speech is related
to the fundamental
frequency (perceived as pitch) of the voice.
35
system function of the form:
H(z) = G
1 −
_p
k=1
akz
−k
. (4.2)
The coefficients G and ak in (4.2) change with time, and should
therefore
be indexed with the subscript ˆn as in (4.1) and Figure 4.1, but
this complication of notation is not generally necessary since it is
usually
clear that the system function or difference equation only applies
over a short time interval.2 Although the linear system is assumed
to
model the composite spectrum effects of radiation, vocal tract tube,
and glottal excitation pulse shape (for voiced speech only) over a
short
time interval, the linear system in the model is commonly referred
to as simply the “vocal tract” system and the corresponding
impulse
response is called the “vocal tract impulse response.” For allpole
linear
systems, as represented by (4.2), the input and output are related by
a difference equation of the form:
s[n] =
_p
k=1
aks[n − k] + Ge[n], (4.3)
where as discussed above, we have suppressed the indication of the
time
at which the difference equation applies.
With such a model as the basis, common practice is to partition
the analysis of the speech signal into techniques for extracting
the parameters of the excitation model, such as the pitch period and
voiced/unvoiced classification, and techniques for extracting the
linear
system model (which imparts the spectrum envelope or spectrum
shaping).
Because of the slowly varying nature of the speech signal, it is
common
to process speech in blocks (also called “frames”) over which the
properties of the speech waveform can be assumed to remain
relatively
constant. This leads to the basic principle of shorttime analysis,
which
2 In general, we use n and m for discrete indices for sequences, but
whenever we want to
indicate a specific analysis time, we use ˆn.
36 ShortTime Analysis of Speech
is represented in a general form by the equation:
Xˆn =
∞_
m=−∞
T {x[m]w[ˆn − m]}, (4.4)
where Xˆn represents the shorttime analysis parameter (or vector
of
parameters) at analysis time ˆn.3 The operator T{ } defines the
nature
of the shorttime analysis function, and w[ˆn − m] represents a
timeshifted
window sequence, whose purpose is to select a segment of the
sequence x[m] in the neighborhood of sample m = ˆn.We will see
several
examples of such operators that are designed to extract or highlight
certain features of the speech signal. The infinite limits in (4.4)
imply
summation over all nonzero values of the windowed segment
xˆn[m] =
x[m]w[ˆn − m]; i.e., for all m in the region of support of the
window.
For example, a finiteduration4 window might be a Hamming
window
defined by
wH[m] =
_
0.54 + 0.46cos(πm/M) −M ≤ m ≤ M
0 otherwise.
(4.5)
Figure 4.2 shows a discretetime Hamming window and its
discretetime
Fourier transform, as used throughout this chapter.5 It can be
shown
that a (2M + 1) sample Hamming window has a frequency main
lobe
(full) bandwidth of 4π/M. Other windows will have similar
properties,
i.e., they will be concentrated in time and frequency, and the
frequency
width will be inversely proportional to the time width [89].
Figure 4.3 shows a 125 ms segment of a speech waveform that
includes both unvoiced (0–50 ms) and voiced speech (50–125 ms).
Also
shown is a sequence of data windows of duration 40 ms and shifted
by
15 ms (320 samples at 16 kHz sampling rate) between windows.
This
illustrates how shorttime analysis is implemented.
3 Sometimes (4.4) is normalized by dividing by the effective
window length, i.e.,
_
∞
m=−∞w[m], so that Xˆn is a weighted average.
4An example of an infiniteduration window is w[n] = nan for n ≥
0. Such a window can
lead to recursive implementations of shorttime analysis functions
[9].
5We have assumed the time origin at the center of a symmetric
interval of 2M + 1 samples.
A causal window would be shifted to the right by M samples.
4.1 ShortTime Energy and ZeroCrossing Rate 37
(a)
(b)
Fig. 4.2 Hamming window (a) and its discretetime Fourier
transform (b).
0 20 40 60 80 100 120
time in ms
Fig. 4.3 Section of speech waveform with shorttime analysis
windows.
4.1 ShortTime Energy and ZeroCrossing Rate
Two basic shorttime analysis functions useful for speech signals
are the
shorttime energy and the shorttime zerocrossing rate. These
functions
are simple to compute, and they are useful for estimating
properties
of the excitation function in the model.
The shorttime energy is defined as
Eˆn =
∞_
m=−∞
(x[m]w[ˆn − m])2 =
∞_
m=−∞
x2[m]w2[ˆn − m]. (4.6)
38 ShortTime Analysis of Speech
In this case the operator T{ } is simply squaring the windowed
samples.
As shown in (4.6), it is often possible to express shorttime
analysis
operators as a convolution or linear filtering operation. In this case,
Eˆn = x2[n] ∗ he[n]
__
n=ˆn, where the impulse response of the linear filter
is he[n] = w2[n].
Similarly, the shorttime zero crossing rate is defined as the
weighted
average of the number of times the speech signal changes sign
within
the time window. Representing this operator in terms of linear
filtering
leads to
Zˆn =
∞_
m=−∞
0.5sgn{x[m]} − sgn{x[m − 1]}w[ˆn − m], (4.7)
where
sgn{x} =
_
1 x ≥ 0
−1 x < 0.
(4.8)
Since 0.5sgn{x[m]} − sgn{x[m − 1]} is equal to 1 if x[m] and x[m
− 1]
have different algebraic signs and 0 if they have the same sign, it
follows
that Zˆn in (4.7) is a weighted sum of all the instances of
alternating sign
(zerocrossing) that fall within the support region of the shifted
window
w[ˆn − m]. While this is a convenient representation that fits the
general
framework of (4.4), the computation of Zˆn could be implemented
in
other ways.
Figure 4.4 shows an example of the shorttime energy and
zerocrossing
rate for a segment of speech with a transition from unvoiced
to voiced speech. In both cases, the window is a Hamming window
(two examples shown) of duration 25 ms (equivalent to 401
samples
at a 16 kHz sampling rate).6 Thus, both the shorttime energy and
the shorttime zerocrossing rate are output of a lowpass filter
whose
frequency response is as shown in Figure 4.2(b). For the 401point
Hamming window used in Figure 4.4, the frequency response is
very
small for discretetime frequencies above 2π/200 rad/s (equivalent
to
16000/200 = 80Hz analog frequency). This means that the short
time
6 In the case of the shorttime energy, the window applied to the
signal samples was the
squareroot of the Hamming window, so that he[n] = w2[n] is the
Hamming window
defined by (4.5).
4.1 ShortTime Energy and ZeroCrossing Rate 39
0 20 40 60 80 100 120
time in ms
Fig. 4.4 Section of speech waveform with shorttime energy and
zerocrossing rate
superimposed.
energy and zerocrossing rate functions are slowly varying
compared
to the time variations of the speech signal, and therefore, they can
be
sampled at a much lower rate than that of the original speech
signal.
For finitelength windows like the Hamming window, this
reduction of
the sampling rate is accomplished by moving the window position
ˆn in
jumps of more than one sample as shown in Figure 4.3.
Note that during the unvoiced interval, the zerocrossing rate is
relatively
high compared to the zerocrossing rate in the voiced interval.
Conversely, the energy is relatively low in the unvoiced region
compared
to the energy in the voiced region. Note also that there is a small
shift of the two curves relative to events in the time waveform.
This is
due to the time delay of M samples (equivalent to 12.5 ms) added
to
make the analysis window filter causal.
The shorttime energy and shorttime zerocrossing rate are
important
because they abstract valuable information about the speech signal,
and they are simple to compute. The shorttime energy is an
indication
of the amplitude of the signal in the interval around time ˆn. From
our model, we expect unvoiced regions to have lower shorttime
energy
than voiced regions. Similarly, the shorttime zerocrossing rate is
a
crude frequency analyzer. Voiced signals have a high frequency
(HF)
falloff due to the lowpass nature of the glottal pulses, while
unvoiced
sounds have much more HF energy. Thus, the shorttime energy
and
shorttime zerocrossing rate can be the basis for an algorithm for
making
a decision as to whether the speech signal is voiced or unvoiced at
40 ShortTime Analysis of Speech
any particular time ˆn. A complete algorithm would involve
measurements
of the statistical distributions of the energy and zerocrossing
rate for both voiced and unvoiced speech segments (and also
background
noise distributions). These distributions can be used to derive
thresholds used in voiced/unvoiced decision [100].
4.2 ShortTime Autocorrelation Function (STACF)
The autocorrelation function is often used as a means of detecting
periodicity in signals, and it is also the basis for many spectrum
analysis
methods. This makes it a useful tool for shorttime speech analysis.
The
STACF is defined as the deterministic autocorrelation function of
the
sequence xˆn[m] = x[m]w[ˆn − m] that is selected by the window
shifted
to time ˆn, i.e.,
φˆn[_] =
∞_
m=−∞
xˆn[m]xˆn[m + _]
=
∞_
m=−∞
x[m]w[ˆn − m]x[m + _]w[ˆn − m − _]. (4.9)
Using the familiar evensymmetric property of the autocorrelation,
φˆn[−_] = φˆn[_], (4.9) can be expressed in terms of linear time
invariant
(LTI) filtering as
φˆn[_] =
∞_
m=−∞
x[m]x[m − _] ˜ w_[ˆn − m], (4.10)
where ˜ w_[m] = w[m]w[m + _]. Note that the STACF is a
twodimensional
function of the discretetime index ˆn (the window position)
and the discretelag index _. If the window has finite duration,
(4.9) can
be evaluated directly or using FFT techniques (see Section 4.6).
For
infiniteduration decaying exponential windows, the shorttime
autocorrelation
of (4.10) can be computed recursively at time ˆn by using a
different filter ˜ w_[m] for each lag value _ [8, 9].
To see how the shorttime autocorrelation can be used in speech
analysis, assume that a segment of the sampled speech signal is a
segment
of the output of the discretetime model shown in Figure 4.1 where
4.2 ShortTime Autocorrelation Function (STACF) 41
the system is characterized at a particular analysis time by an
impulse
response h[n], and the input is either a periodic impulse train or
random
white noise. (Note that we have suppressed the indication of
analysis
time.) Different segments of the speech signal will have the same
form
of model with different excitation and system impulse response.
That
is, assume that s[n] = e[n] ∗ h[n], where e[n] is the excitation to
the
linear system with impulse response h[n]. A well known, and
easily
proved, property of the autocorrelation function is that
φ(s)[_] = φ(e)[_] ∗ φ(h)[_], (4.11)
i.e., the autocorrelation function of s[n] = e[n] ∗ h[n] is the
convolution
of the autocorrelation functions of e[n] and h[n]. In the case of the
speech signal, h[n] represents the combined (by convolution)
effects
of the glottal pulse shape (for voiced speech), vocal tract shape,
and
radiation at the lips. For voiced speech, the autocorrelation of a
periodic
impulse train excitation with period P0 is a periodic impulse train
sequence with the same period. In this case, therefore, the
autocorrelation
of voiced speech is the periodic autocorrelation function
φ(s)[_] =
∞_
m=−∞
φ(h)[_ − mP0]. (4.12)
In the case of unvoiced speech, the excitation can be assumed to be
random
white noise, whose stochastic autocorrelation function would be
an impulse sequence at _ = 0. Therefore, the autocorrelation
function
of unvoiced speech computed using averaging would be simply
φ(s)[_] = φ(h)[_]. (4.13)
Equation (4.12) assumes periodic computation of an infinite
periodic
signal, and (4.13) assumes probability average or averaging over
an
infinite time interval (for stationary random signals). However, the
deterministic autocorrelation function of a finitelength segment of
the
speech waveform will have properties similar to those of (4.12)
and
(4.13) except that the correlation values will taper off with lag _
due
to the tapering of the window and the fact that less and less data is
involved in the computation of the shorttime autocorrelation for
longer
42 ShortTime Analysis of Speech
0 10 20 30 40
−0.2
−0.1
0
0.1
0.2
0 10 20 30 40
−0.2
−0.1
0
0.1
0.2
0 10 20 30 40
−0.5
0
0.5
1
0 20 40
−0.5
0
0.5
1
(a) Voiced Segment (b) Voiced Autocorrelation
(c) Unvoiced Segment (d) Unvoiced Autocorrelation
time in ms time in ms
Fig. 4.5 Voiced and unvoiced segments of speech and their
corresponding STACF.
lag values. This tapering off in level is depicted in Figure 4.5 for
both a
voiced and an unvoiced speech segment. Note the peak in the
autocorrelation
function for the voiced segment at the pitch period and twice
the pitch period, and note the absence of such peaks in the
autocorrelation
function for the unvoiced segment. This suggests that the STACF
could be the basis for an algorithm for estimating/detecting the
pitch
period of speech. Usually such algorithms involve the
autocorrelation
function and other shorttime measurements such as zerocrossings
and
energy to aid in making the voiced/unvoiced decision.
Finally, observe that the STACF implicitly contains the shorttime
energy since
Eˆn =
∞_
m=−∞
(x[m]w[ˆn − m])2 = φˆn[0]. (4.14)
4.3 ShortTime Fourier Transform (STFT)
The shorttime analysis functions discussed so far are examples of
the general shorttime analysis principle that is the basis for most
4.3 ShortTime Fourier Transform (STFT) 43
algorithms for speech processing. We now turn our attention to
what is
perhaps the most important basic concept in digital speech
processing.
In subsequent chapters, we will find that the STFT defined as
Xˆn(ej ˆω) =
∞_
m=−∞
x[m]w[ˆn − m]e
−j ˆωm, (4.15)
is the basis for a wide range of speech analysis, coding and
synthesis systems.
By definition, for fixed analysis time ˆn, the STFT is the
discretetime
Fourier transform (DTFT) of the signal xˆn[m] = x[m]w[ˆn − m],
i.e., the DTFT of the (usually finiteduration) signal selected and
amplitudeweighted by the sliding window w[ˆn − m] [24, 89,
129]. Thus,
the STFT is a function of two variables; ˆn the discretetime index
denoting
the window position, and ˆω representing the analysis frequency.7
Since (4.15) is a sequence of DTFTs, the twodimensional function
Xˆn(ej ˆω) at discretetime ˆn is a periodic function of continuous
radian
frequency ˆω with period 2π [89].
As in the case of the other shorttime analysis functions discussed
in this chapter, the STFT can be expressed in terms of a linear
filtering
operation. For example, (4.15) can be expressed as the discrete
convolution
Xˆn(ej ˆω) = (x[n]e
−j ˆωn) ∗ w[n]
___
n=ˆn
, (4.16)
or, alternatively,
Xˆn(ej ˆω) =
_
x[n] ∗ (w[n]ej ˆωn)
_
e
−j ˆωn
___
n=ˆn
. (4.17)
Recall that a typical window like a Hamming window, when
viewed
as a linear filter impulse response, has a lowpass frequency
response
with the cutoff frequency varying inversely with the window
length.
(See Figure 4.2(b).) This means that for a fixed value of ˆω, Xˆn(ej
ˆω)
is slowly varying as ˆn varies. Equation (4.16) can be interpreted as
follows: the amplitude modulation x[n]e−j ˆωn shifts the spectrum
of
7 As before, we use ˆn to specify the analysis time, and ˆω is used
to distinguish the STFT
analysis frequency from the frequency variable of the nontime
dependent Fourier transform
frequency variable ω.
44 ShortTime Analysis of Speech
x[n] down by ˆω, and the window (lowpass) filter selects the
resulting
band of frequencies around zero frequency. This is, of course, the
band of frequencies of x[n] that were originally centered on
analysis
frequency ˆω. An identical conclusion follows from (4.17): x[n] is
the
input to a bandpass filter with impulse response w[n]ej ˆωn, which
selects
the band of frequencies centered on ˆω. Then that band of
frequencies is
shifted down by the amplitude modulation with e−j ˆωn, resulting
again
in the same lowpass signal [89].
In summary, the STFT has three interpretations: (1) It is a
sequence
of discretetime Fourier transforms of windowed signal segments,
i.e., a
periodic function of ˆω at each window position ˆn. (2) For each
frequency
ˆω with ˆn varying, it is the time sequence output of a lowpass
filter that
follows frequency downshifting by ˆω. (3) For each frequency ˆω,
it is
the time sequence output resulting from frequency downshifting
the
output of a bandpass filter.
4.4 Sampling the STFT in Time and Frequency
As defined in (4.15), the STFT is a function of a continuous
analysis
frequency ˆω. The STFT becomes a practical tool for both analysis
and applications when it is implemented with a finiteduration
window
moved in steps of R >1 samples in time and computed at a discrete
set of frequencies as in
XrR[k] =
_rR
m=rR−L+1
x[m]w[rR − m]e
−j(2πk/N)m k = 0,1,...,N − 1,
(4.18)
where N is the number of uniformly spaced frequencies across the
interval 0 ≤ ˆω < 2π, and L is the window length (in samples).
Note
that we have assumed that w[m] is causal and nonzero only in the
range 0 ≤ m ≤ L − 1 so that the windowed segment x[m]w[rR − m]
is nonzero over rR − L + 1 ≤ m ≤ rR. To aid in interpretation, it is
helpful to write (4.18) in the equivalent form:
XrR[k] = ˜XrR[k]e
−j(2πk/N)rR k = 0,1,...,N − 1, (4.19)
4.4 Sampling the STFT in Time and Frequency 45
where
˜X
rR[k] =
L_−1
m=0
x[rR − m]w[m]ej(2πk/N)m k = 0,1,...,N − 1. (4.20)
Since we have assumed, for specificity, that w[m] _= 0 only in the
range
0 ≤ m ≤ L − 1, the alternative form, ˜XrR[k], has the interpretation
of
an Npoint DFT of the sequence x[rR − m]w[m], which, due to the
definition of the window, is nonzero in the interval 0 ≤ m ≤ L − 1.8
In
(4.20), the analysis time rR is shifted to the time origin of the DFT
computation, and the segment of the speech signal is the time
reversed
sequence of L samples that precedes the analysis time. The
complex
exponential factor e−j(2πk/N)rR in (4.19) results from the shift of
the
time origin.
For a discretetime Fourier transform interpretation, it follows that
˜X
rR[k] can be computed by the following process:
(1) Form the sequence xrR[m] = x[rR − m]w[m], for m =
0,1,...,L − 1.
(2) Compute the complex conjugate of the Npoint DFT of
the sequence xrR[m]. (This can be done efficiently with an
Npoint FFT algorithm.)
(3) The multiplication by e−j(2πk/N)rR can be done if necessary,
but often can be omitted (as in computing a sound spectrogram
or spectrographic display).
(4) Move the time origin by R samples (i.e., r → r + 1) and
repeat steps (1), (2), and (3), etc.
The remaining issue for complete specification of the sampled
STFT
is specification of the temporal sampling period, R, and the number
of
uniform frequencies, N. It can easily be shown that both R and N
are determined entirely by the time width and frequency bandwidth
of
the lowpass window, w[m], used to compute the STFT [1], giving
the
8 The DFT is normally defined with a negative exponent. Thus,
since x[rR − m]w[m] is
real, (4.20) is the complex conjugate of the DFT of the windowed
sequence [89].
46 ShortTime Analysis of Speech
following constraints on R and N:
(1) R ≤ L/(2C) where C is a constant that is dependent on the
window frequency bandwidth; C = 2 for a Hamming window,
C = 1 for a rectangular window.
(2) N ≥ L, where L is the window length in samples.
Constraint (1) above is related to sampling the STFT in time at a
rate of twice the window bandwidth in frequency in order to
eliminate
aliasing in frequency of the STFT, and constraint (2) above is
related
to sampling in frequency at a rate of twice the equivalent time
width
of the window to ensure that there is no aliasing in time of the
STFT.
4.5 The Speech Spectrogram
Since the 1940s, the sound spectrogram has been a basic tool for
gaining
understanding of how the sounds of speech are produced and how
phonetic information is encoded in the speech signal. Up until the
1970s, spectrograms were made by an ingenious device comprised
of
an audio tape loop, variable analog bandpass filter, and electrically
sensitive paper [66]. Today spectrograms like those in Figure 4.6
are
made by DSP techniques [87] and displayed as either pseudocolor
or
grayscale images on computer screens.
Sound spectrograms like those in Figure 4.6 are simply a display of
the magnitude of the STFT. Specifically, the images in Figure 4.6
are
plots of
S(tr,fk) = 20log10 ˜XrR[k] = 20log10 XrR[k], (4.21)
where the plot axes are labeled in terms of analog time and
frequency
through the relations tr = rRT and fk = k/(NT), where T is
the sampling period of the discretetime signal x[n] = xa(nT). In
order
to make smooth looking plots like those in Figure 4.6, R is usually
quite small compared to both the window length L and the number
of samples in the frequency dimension, N, which may be much
larger than the window length L. Such a function of two variables
can be plotted on a two dimensional surface (such as this text) as
either a grayscale or a colormapped image. Figure 4.6 shows the
time
4.5 The Speech Spectrogram 47
−1
0
1 SH UH D W IY CH EY S
frequency in kHz
0 100 200 300 400 500 600
0
1
2
3
4
5
−100
−50
0
frequency in kHz
time in ms
0 100 200 300 400 500 600
0
1
2
3
4
5
−120
−100
−80
−60
−40
−20
0
Fig. 4.6 Spectrogram for speech signal of Figure 1.1.
waveform at the top and two spectrograms computed with different
length analysis windows. The bars on the right calibrate the color
map (in dB).
A careful interpretation of (4.20) and the corresponding
spectrogram
images leads to valuable insight into the nature of the speech
signal. First note that the window sequence w[m] is nonzero only
over
the interval 0 ≤ m ≤ L − 1. The length of the window has a major
effect on the spectrogram image. The upper spectrogram in Figure
4.6
was computed with a window length of 101 samples,
corresponding to
10 ms time duration. This window length is on the order of the
length
of a pitch period of the waveform during voiced intervals. As a
result,
in voiced intervals, the spectrogram displays vertically oriented
striations
corresponding to the fact that the sliding window sometimes
includes mostly large amplitude samples, then mostly small
amplitude
48 ShortTime Analysis of Speech
samples, etc. As a result of the short analysis window, each
individual
pitch period is resolved in the time dimension, but the resolution
in the frequency dimension is poor. For this reason, if the analysis
window is short, the spectrogram is called a wideband
spectrogram.
This is consistent with the linear filtering interpretation of the
STFT,
since a short analysis filter has a wide passband. Conversely, when
the
window length is long, the spectrogram is a narrowband
spectrogram,
which is characterized by good frequency resolution and poor time
resolution.
The upper plot in Figure 4.7, for example, shows S(tr,fk) as a
function of fk at time tr = 430 ms. This vertical slice through the
spectrogram is at the position of the black vertical line in the upper
spectrogram of Figure 4.6. Note the three broad peaks in the
spectrum
slice at time tr = 430 ms, and observe that similar slices would
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−80
−60
−40
−20
0
log magnitude (dB)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
−80
−60
−40
−20
0
frequency in kHz
log magnitude (dB)
Fig. 4.7 Shorttime spectrum at time 430 ms (dark vertical line in
Figure 4.6) with Hamming
window of length M = 101 in upper plot and M = 401 in lower
plot.
4.6 Relation of STFT to STACF 49
be obtained at other times around tr = 430 ms. These large peaks
are
representative of the underlying resonances of the vocal tract at the
corresponding time in the production of the speech signal.
The lower spectrogram in Figure 4.6 was computed with a window
length of 401 samples, corresponding to 40 ms time duration. This
window length is on the order of several pitch periods of the
waveform
during voiced intervals. As a result, the spectrogram no longer
displays
vertically oriented striations since several periods are included in
the
window no matter where it is placed on the waveform in the
vicinity
of the analysis time tr. As a result, the spectrogram is not as
sensitive
to rapid time variations, but the resolution in the frequency
dimension
is much better. Therefore, the striations tend to be horizontally
oriented
in the narrowband case since the fundamental frequency and its
harmonics are all resolved. The lower plot in Figure 4.7, for
example,
shows S(tr,fk) as a function of fk at time tr = 430 ms. In this case,
the signal is voiced at the position of the window, so over the
analysis
window interval it acts very much like a periodic signal. Periodic
signals have Fourier spectra that are composed of impulses at the
fundamental
frequency, f0, and at integer multiples (harmonics) of the
fundamental frequency [89]. Multiplying by the analysis window
in the
timedomain results in convolution of the Fourier transform of the
window
with the impulses in the spectrum of the periodic signal [89]. This
is evident in the lower plot of Figure 4.7, where the local maxima
of the
curve are spaced at multiples of the fundamental frequency f0 =
1/T0,
where T0 is the fundamental period (pitch period) of the signal.9
4.6 Relation of STFT to STACF
A basic property of Fourier transforms is that the inverse Fourier
transform
of the magnitudesquared of the Fourier transform of a signal is the
autocorrelation function for that signal [89]. Since the STFT
defined in
(4.15) is a discretetime Fourier transform for fixed window
position,
it follows that φˆn[_] given by (4.9) is related to the STFT given by
9 Each “ripple” in the lower plot is essentially a frequencyshifted
copy of the Fourier transform
of the Hamming window used in the analysis.
50 ShortTime Analysis of Speech
(4.15) as
φˆn[_] =
1
2π
_ π
−π
Xˆn(ej ˆω)2ej ˆω_dˆω. (4.22)
The STACF can also be computed from the sampled STFT. In
particular,
an inverse DFT can be used to compute
˜φrR[_] =
1
N
N_−1
k=0
˜XrR(ej(2πk/N))2ej(2πk/N)_ (4.23)
and ˜φrR[_] = φrR[_] for _ = 0,1,...,L − 1 if N ≥ 2L. If L<N <2L
time aliasing will occur, but ˜φrR[_] = φrR[_] for _ = 0,1,...,N − L.
Note
that (4.14) shows that the shorttime energy can also be obtained
from
either (4.22) or (4.23) by setting _ = 0.
Figure 4.8 illustrates the equivalence between the STACF and the
STFT. Figures 4.8(a) and 4.8(c) show the voiced and unvoiced
autocorrelation
functions that were shown in Figures 4.5(b) and 4.5(d). On the
right are the corresponding STFTs. The peaks around 9 and 18 ms
in
0 10 20 30 40
−0.5
0
0.5
1
0 10 20 30 40
−0.5
0
0.5
1
0 2000 4000 6000 8000
−100
−50
0
50
0 2000 4000 6000 8000
−100
−50
0
50
(a) Voiced Autocorrelation (b) Voiced STFT
(c) Unvoiced Autocorrelation (d) Unvoiced STFT
time in msec frequency in kHz
Fig. 4.8 STACF and corresponding STFT.
4.7 ShortTime Fourier Synthesis 51
the autocorrelation function of the voiced segment imply a
fundamental
frequency at this time of approximately 1000/9 = 111 Hz.
Similarly,
note in the STFT on the right in Figure 4.8(b) that there are
approximately
18 local regularly spaced peaks in the range 0–2000 Hz. Thus,
we can estimate the fundamental frequency to be 2000/18 = 111
Hz as
before.
4.7 ShortTime Fourier Synthesis
From the linear filtering point of view of the STFT, (4.19) and
(4.20)
represent the downsampled (by the factor R) output of the process
of bandpass filtering with impulse response w[n]ej(2πk/N)n
followed
by frequencydownshifting by (2πk/N). Alternatively, (4.18)
shows
that XrR[k] is a downsampled output of the lowpass window filter
with frequencydownshifted input x[n]e−j(2πk/N)n. The left half
of Figure
4.9 shows a block diagram representation of the STFT as a
combination
of modulation, followed by lowpass filtering, followed by a
downsampler by R. This structure is often called a filter bank, and
the outputs of the individual filters are called the channel signals.
The
w[n]
w[n]
w[n]
x
x
x x
x
x
Short Time Modifications
R
R
R
R
R
R
f [n]
f [n]
f [n]
+
Xn[k]
x[n]
j N n
e N
( 1) 2 − π −
j N n
e N
( 1) 2 π −
j n
e N
(1) − 2π
j n
e N
(0) − 2π
j n
e N
(1) 2π
j n
e N
(0) 2π
y[n]
XrR[k] YrR[k]
Fig. 4.9 Filterbank interpretation of shorttime Fourier analysis and
synthesis.
52 ShortTime Analysis of Speech
spectrogram is comprised of the set of outputs of the filter bank
with
each channel signal corresponding to a horizontal line in the
spectrogram.
If we want to use the STFT for other types of speech processing,
we may need to consider if and how it is possible to reconstruct
the speech signal from the STFT, i.e., from the channel signals.
The
remainder of Figure 4.9 shows how this can be done.
First, the diagram shows the possibility that the STFT might be
modified by some sort of processing. An example might be
quantization
of the channel signals for data compression. The modified STFT is
denoted as YrR[k]. The remaining parts of the structure on the right
in
Figure 4.9 implement the synthesis of a new time sequence y[n]
from
the STFT. This part of the diagram represents the synthesis
equation
y[n] =
N_−1
k=0
_ ∞_
r=−∞
YrR[k]f[n − rR]
ej(2πk/N)n. (4.24)
The steps represented by (4.24) and the right half of Figure 4.9
involve
first upsampling by R followed by linear filtering with f[n] (called
the synthesis window). This operation, defined by the part within
the
parenthesis in (4.24), interpolates the STFT YrR[k] to the time
sampling
rate of the speech signal for each channel k.10 Then, synthesis
is achieved by modulation with ej(2πk/N)n, which upshifts the
lowpass
interpolated channel signals back to their original frequency bands
centered
on frequencies (2πk/N). The sum of the upshifted signals is the
synthesized output.
There are many variations on the shorttime Fourier analysis/
synthesis paradigm. The FFT algorithm can be used to implement
both analysis and synthesis; special efficiencies result when L > R
= N,
and the analysis and synthesis can be accomplished using only real
filtering
and with modulation operations being implicitly achieved by
the downsampling [24, 129]. However, it is most important to note
that with careful choice of the parameters L, R, and N and careful
design of the analysis window w[m] together with the synthesis
window
f[n], it is possible to reconstruct the signal x[n] with negligible
10 The sequence f[n] is the impulse response of a LTI filter with
gain R/N and normalized
cutoff frequency π/R. It is often a scaled version of w[n].
4.8 ShortTime Analysis is Fundamental to our Thinking 53
error (y[n] = x[n]) from the unmodified STFT YrR[k] = XrR[k].
One
condition that guarantees exact reconstruction is [24, 129].
∞_
r=−∞
w[rR − n + qN]f[n − rR] =
_
1 q = 0
0 q _= 0.
(4.25)
Shorttime Fourier analysis and synthesis is generally formulated
with equally spaced channels with equal bandwidths. However, we
have seen that models for auditory processing involve nonuniform
filter
banks. Such filter banks can be implemented as tree structures
where
frequency bands are successively divided into low and HF bands.
Such
structures are essentially the same as wavelet decompositions, a
topic
beyond the scope of this text, but one for which a large body of
literature
exists. See, for example [18, 125, 129].
4.8 ShortTime Analysis is Fundamental to our Thinking
It can be argued that the shorttime analysis principle, and
particularly
the shorttime Fourier representation of speech, is fundamental to
our thinking about the speech signal and it leads us to a wide
variety
of techniques for achieving our goal of moving from the sampled
time
waveform back along the speech chain toward the implicit
message.
The fact that almost perfect reconstruction can be achieved from
the
filter bank channel signals gives the shorttime Fourier
representation
major credibility in the digital speech processing tool kit. This
importance
is strengthened by the fact that, as discussed briefly in Chapter 3,
models for auditory processing are based on a filter bank as the
first
stage of processing. Much of our knowledge of perceptual effects
is
framed in terms of frequency analysis, and thus, the STFT
representation
provides a natural framework within which this knowledge can be
represented and exploited to obtain efficient representations of
speech
and more general audio signals. We will have more to say about
this
in Chapter 7, but first we will consider (in Chapter 5) the technique
of cepstrum analysis, which is based directly on the STFT, and the
technique of linear predictive analysis (in Chapter 6), which is
based
on the STACF and thus equivalently on the STFT.
5
Homomorphic Speech Analysis
The STFT provides a useful framework for thinking about almost
all
the important techniques for analysis of speech that have been
developed
so far. An important concept that flows directly from the STFT
is the cepstrum, more specifically, the shorttime cepstrum, of
speech.
In this chapter, we explore the use of the shorttime cepstrum as a
representation of speech and as a basis for estimating the
parameters
of the speech generation model. A more detailed discussion of the
uses
of the cepstrum in speech processing can be found in [110].
5.1 Definition of the Cepstrum and Complex Cepstrum
The cepstrum was defined by Bogert, Healy, and Tukey to be the
inverse
Fourier transform of the log magnitude spectrum of a signal [16].
Their
original definition, loosely framed in terms of spectrum analysis of
analog
signals, was motivated by the fact that the logarithm of the Fourier
spectrum of a signal containing an echo has an additive periodic
component
depending only on the echo size and delay, and that further
Fourier analysis of the log spectrum can aid in detecting the
presence
54
5.1 Definition of the Cepstrum and Complex Cepstrum 55
of that echo. Oppenheim, Schafer, and Stockham showed that the
cepstrum
is related to the more general concept of homomorphic filtering
of signals that are combined by convolution [85, 90, 109]. They
gave a
definition of the cepstrum of a discretetime signal as
c[n] =
1
2π
_ π
−π
log X(ejω)ejωndω, (5.1a)
where log X(ejω) is the logarithm of the magnitude of the DTFT
of the signal, and they extended the concept by defining the
complex
cepstrum as
ˆx[n] =
1
2π
_ π
−π
log{X(ejω)}ejωndω, (5.1b)
where log{X(ejω)} is the complex logarithm of X(ejω) defined as
ˆX
(ejω) = log{X(ejω)} = log X(ejω) + j arg[X(ejω)]. (5.1c)
The transformation implied by (5.1b) is depicted as the block
diagram
in Figure 5.1. The same diagram represents the cepstrum if the
complex
logarithm is replaced by the logarithm of the magnitude of the
DTFT.
Since we restrict our attention to real sequences, x[n], it follows
from
the symmetry properties of Fourier transforms that the cepstrum is
the
even part of the complex cepstrum, i.e., c[n] = (ˆx[n] + ˆx[−n])/2
[89].
As shown in Figure 5.1, the operation of computing the complex
cepstrum
from the input can be denoted as ˆx[n] = D∗{x[n]}. In the theory
of homomorphic systems D∗{ } is called the characteristic system
for
convolution. The connection between the cepstrum concept and
homomorphic
filtering of convolved signals is that the complex cepstrum has
the property that if x[n] = x1[n] ∗ x2[n], then
ˆx[n] = D∗{x1[n] ∗ x2[n]} = ˆx1[n] + ˆx2[n]. (5.2)
Fig. 5.1 Computing the complex cepstrum using the DTFT.
56 Homomorphic Speech Analysis
Fig. 5.2 The inverse of the characteristic system for convolution
(inverse complex cepstrum).
That is, the complex cepstrum operator transforms convolution
into
addition. This property, which is true for both the cepstrum and the
complex cepstrum, is what makes the cepstrum and the complex
cepstrum
useful for speech analysis, since our model for speech production
involves convolution of the excitation with the vocal tract impulse
response, and our goal is often to separate the excitation signal
from
the vocal tract signal. In the case of the complex cepstrum, the
inverse
of the characteristic system exists as in Figure 5.2, which shows
the
reverse cascade of the inverses of the operators in Figure 5.1.
Homomorphic
filtering of convolved signals is achieved by forming a modified
complex cepstrum
ˆy[n] = g[n]ˆx[n], (5.3)
where g[n] is a window (a “lifter” in the terminology of Bogert et
al.)
which selects a portion of the complex cepstrum for inverse
processing.
A modified output signal y[n] can then be obtained as the output of
Figure 5.2 with ˆy[n] given by (5.3) as input. Observe that (5.3)
defines
a linear operator in the conventional sense, i.e., if ˆx[n] = ˆx1[n] +
ˆx2[n]
then ˆy[n] = g[n]ˆx1[n] + g[n]ˆx2[n]. Therefore, the output of the
inverse
characteristic system will have the form y[n] = y1[n] ∗ y2[n],
where
ˆy1[n] = g[n]ˆx1[n] is the complex cepstrum of y1[n], etc.
Examples of
lifters used in homomorphic filtering of speech are given in
Sections 5.4
and 5.6.2.
The key issue in the definition and computation of the complex
cepstrum
is the computation of the complex logarithm; more specifically,
the computation of the phase angle arg[X(ejω)], which must be
done
so as to preserve an additive combination of phases for two signals
combined by convolution [90, 109].
5.2 The ShortTime Cepstrum 57
The independent variable of the cepstrum and complex cepstrum
is nominally time. The crucial observation leading to the cepstrum
concept is that the log spectrum can be treated as a waveform to be
subjected to further Fourier analysis. To emphasize this
interchanging
of domains of reference, Bogert et al. [16] coined the word
cepstrum
by transposing some of the letters in the word spectrum. They
created
many other special terms in this way including quefrency as the
name
for the independent variable of the cepstrum and liftering for the
operation
of linear filtering the log magnitude spectrum by the operation
of (5.3). Only the terms cepstrum, quefrency, and liftering are
widely
used today.
5.2 The ShortTime Cepstrum
The application of these definitions to speech requires that the
DTFT be replaced by the STFT. Thus the shorttime cepstrum is
defined as
cˆn[m] =
1
2π
_ π
−π
log Xˆn(ej ˆω)ej ˆωmdˆω, (5.4)
where Xˆn(ej ˆω) is the STFT defined in (4.15), and the shorttime
complex cepstrum is likewise defined by replacing X(ej ˆω) by
Xˆn(ej ˆω)
in (5.1b).1 The similarity to the STACF defined by (4.9) should be
clear. The shorttime cepstrum is a sequence of cepstra of
windowed
finiteduration segments of the speech waveform. By analogy, a
“cepstrogram”
would be an image obtained by plotting the magnitude of
the shorttime cepstrum as a function of quefrency m and analysis
time ˆn.
5.3 Computation of the Cepstrum
In (5.1a) and (5.1b), the cepstrum and complex cepstrum are
defined
in terms of the DTFT. This is useful in the basic definitions, but
not
1 In cases where we wish to explicitly indicate analysistime
dependence of the shorttime
cepstrum, we will use ˆn for the analysis time and m for quefrency
as in (5.4), but as in
other instances, we often suppress the subscript ˆn.
58 Homomorphic Speech Analysis
for use in processing sampled speech signals. Fortunately, several
computational
options exist.
5.3.1 Computation Using the DFT
Since the DFT (computed with an FFT algorithm) is a sampled
(in frequency) version of the DTFT of a finitelength sequence
(i.e.,
X[k] = X(ej2πk/N) [89]), the DFT and inverse DFT can be
substituted
for the DTFT and its inverse in (5.1a) and (5.1b) as shown in
Figure 5.3, which shows that the complex cepstrum can be
computed
approximately using the equations
X[k] =
N_−1
n=0
x[n]e
−j(2πk/N)n (5.5a)
ˆX
[k] = log X[k] + j arg{X[k]} (5.5b)
˜ˆ
x[n] =
1
N
N_−1
k=0
ˆX
[k]ej(2πk/N)n. (5.5c)
Note the “tilde” ˜ symbol above ˆx[n] in (5.5c) and in Figure 5.3.
Its
purpose is to emphasize that using the DFT instead of the DTFT
results in an approximation to (5.1b) due to the timedomain
aliasing
resulting from the sampling of the log of the DTFT [89]. That is,
˜ˆ
x[n] =
∞_
r=−∞
ˆx[n + rN], (5.6)
where ˆx[n] is the complex cepstrum defined by (5.1b). An
identical
equation holds for the timealiased cepstrum ˜c[n].
The effect of timedomain aliasing can be made negligible by
using a
large value for N. A more serious problem in computation of the
complex
cepstrum is the computation of the complex logarithm. This is
Fig. 5.3 Computing the cepstrum or complex cepstrum using the
DFT.
5.3 Computation of the Cepstrum 59
because the angle of a complex number is usually specified
modulo
2π, i.e., by the principal value. In order for the complex cepstrum
to be evaluated properly, the phase of the sampled DTFT must be
evaluated as samples of a continuous function of frequency. If the
principal
value phase is first computed at discrete frequencies, then it must
be
“unwrapped” modulo 2π in order to ensure that convolutions are
transformed
into additions. A variety of algorithms have been developed for
phase unwrapping [90, 109, 127]. While accurate phase
unwrapping
presents a challenge in computing the complex cepstrum, it is not a
problem in computing the cepstrum, since the phase is not used.
Furthermore,
phase unwrapping can be avoided by using numerical computation
of the poles and zeros of the ztransform.
5.3.2 zTransform Analysis
The characteristic system for convolution can also be represented
by
the twosided ztransform as depicted in Figure 5.4. This is very
useful
for theoretical investigations, and recent developments in
polynomial
root finding [117] have made the ztransform representation a
viable
computational basis as well. For this purpose, we assume that the
input
signal x[n] has a rational ztransform of the form:
X(z) = Xmax(z)Xuc(z)Xmin(z), (5.7)
where
Xmax(z) = zMo
_Mo
k=1
(1 − akz
−1) =
_Mo
k=1
(−ak)
_Mo
k=1
(1 − a
−1
k z), (5.8a)
Xuc(z) =
M_uc
k=1
(1 − ejθkz
−1), (5.8b)
Xmin(z) = A
_Mi
k=1
(1 − bkz
−1)
_Ni
k=1
(1 − ckz
−1)
. (5.8c)
The zeros of Xmax(z), i.e., zk = ak, are zeros of X(z) outside of the
unit circle (ak > 1). Xmax(z) is thus the maximumphase part of
X(z).
60 Homomorphic Speech Analysis
Fig. 5.4 ztransform representation of characteristic system for
convolution.
Xuc(z) contains all the zeros (with angles θk) on the unit circle.
The
minimumphase part is Xmin(z), where bk and ck are zeros and
poles,
respectively, that are inside the unit circle (bk < 1 and ck < 1).
The
factor zMo implies a shift of Mo samples to the left. It is included
to
simplify the results in (5.11).
The complex cepstrum of x[n] is determined by assuming that the
complex logarithm log{X(z)} results in the sum of logarithms of
each
of the product terms, i.e.,
ˆX
(z) = log
_____
_Mo
k=1
(−ak)
_____
+
_Mo
k=1
log(1 − a
−1
k z) +
M_uc
k=1
log(1 − ejθkz
−1)
+ log A +
_Mi
k=1
log(1 − bkz
−1) −
_Ni
k=1
log(1 − ckz
−1). (5.9)
Applying the power series expansion
log(1 − a) = −
∞_
n=1
an
n
a < 1 (5.10)
to each of the terms in (5.9) and collecting the coefficients of the
positive
and negative powers of z gives
ˆx[n] =
),,,,,,,,,§
,,,,,,,,,§
_Mo
k=1
ank
n
n < 0
log A + log
_____
_mo
k=1
(−ak)
_____
n = 0
_
−
M_uc
k=1
ejθkn
n
−
_Mi
k=1
bnk
n
+
_Ni
k=1
cnk
n
n > 0.
(5.11)
Given all the poles and zeros of a ztransform X(z), (5.11) allows
us to compute the complex cepstrum with no approximation. This
is
5.3 Computation of the Cepstrum 61
the case in theoretical analysis where the poles and zeros are
specified.
However, (5.11) is also useful as the basis for computation. All
that is
needed is a process for obtaining the ztransform as a rational
function
and a process for finding the zeros of the numerator and
denominator.
This has become more feasible with increasing computational
power
and with new advances in finding roots of large polynomials [117].
One method of obtaining a ztransform is simply to select a
finitelength
sequence of samples of a signal. The ztransform is then simply
a polynomial with the samples x[n] as coefficients, i.e.,
X(z) =
M_
n=0
x[n]z
−n = A
_Mo
k=1
(1 − akz
−1)
_Mi
k=1
(1 − bkz
−1). (5.12)
A second method that yields a ztransform is the method of linear
predictive analysis, to be discussed in Chapter 6.
5.3.3 Recursive Computation of the Complex Cepstrum
Another approach to computing the complex cepstrum applies only
to
minimumphase signals, i.e., signals having a ztransform whose
poles
and zeros are inside the unit circle. An example would be the
impulse
response of an allpole vocal tract model with system function
H(z) = G
1 −
_p
k=1
αkz
−____________k
= G
_p
k=1
(1 − ckz
−1)
. (5.13)
Such models are implicit in the use of linear predictive analysis of
speech
(Chapter 6). In this case, all the poles ck must be inside the unit
circle
for stability of the system. From (5.11) it follows that the complex
cepstrum of the impulse response h[n] corresponding to H(z) is
ˆh
[n] =
),,,,§
,,,,§
0 n < 0
log G n = 0
_p
k=1
cnk
n
n > 0.
(5.14)
62 Homomorphic Speech Analysis
It can be shown [90] that the impulse response and its complex
cepstrum
are related by the recursion formula:
ˆh
[n] =
),,,,§
,,,,§
0 n < 0
logG n= 0
h[n]
h[0]
−
n_−1
k=1
_
k
n
_ˆh[k]h[n − k]
h[0] n ≥ 1.
(5.15)
Furthermore, working with the reciprocal (negative logarithm) of
(5.13), it can be shown that there is a direct recursive relationship
between the coefficients of the denominator polynomial in (5.13)
and
the complex cepstrum of the impulse response of the model filter,
i.e.,
ˆh
[n] =
),,,,§
,,,,§
0 n < 0
logG n= 0
αn +
n_−1
k=1
_
k
n
_
ˆh
[k]αn−k n > 0.
(5.16)
From (5.16) it follows that the coefficients of the denominator
polynomial
can be obtained from the complex cepstrum through
αn =ˆh[n] −
n_−1
k=1
_
k
n
_
ˆh
[k]αn−k 1 ≤ n ≤ p. (5.17)
From (5.17), it follows that p + 1 values of the complex cepstrum
are
sufficient to fully determine the speech model system in (5.13)
since all
the denominator coefficients and G can be computed from ˆh[n] for
n =
0,1, . . . , p using (5.17). This fact is the basis for the use of the
cepstrum
in speech coding and speech recognition as a vector representation
of
the vocal tract properties of a frame of speech.
5.4 ShortTime Homomorphic Filtering of Speech
Figure 5.5 shows an example of the shorttime cepstrum of speech
for
the segments of voiced and unvoiced speech in Figures 4.5(a) and
4.5(c).
The low quefrency part of the cepstrum is expected to be
representative
of the slow variations (with frequency) in the log spectrum, while
the
5.4 ShortTime Homomorphic Filtering of Speech 63
0 10 20 30 40
−1
0
1
0 10 20 30 40
−1
0
1
0 2000 4000 6000 8000
−8
−6
−4
−2
0
2
4
0 2000 4000 6000 8000
−8
−6
−4
−2
0
2
4
(a) Voiced Cepstrum (b) Voiced log STFT
(c) Unvoiced Cepstrum (d) Unvoiced log STFT
time in ms frequency in kHz
Fig. 5.5 Shorttime cepstra and corresponding STFTs and
homomorphicallysmoothed
spectra.
high quefrency components would correspond to the more rapid
fluctuations
of the log spectrum. This is illustrated by the plots in Figure 5.5.
The log magnitudes of the corresponding shorttime spectra are
shown
on the right as Figures 5.5(b) and 5.5(d). Note that the spectrum for
the voiced segment in Figure 5.5(b) has a structure of periodic
ripples
due to the harmonic structure of the quasiperiodic segment of
voiced
speech. This periodic structure in the log spectrum of Figure
5.5(b),
manifests itself in the cepstrum peak at a quefrency of about 9ms
in Figure 5.5(a). The existence of this peak in the quefrency range
of
expected pitch periods strongly signals voiced speech.
Furthermore, the
quefrency of the peak is an accurate estimate of the pitch period
during
the corresponding speech interval. As shown in Figure 4.5(b), the
autocorrelation function also displays an indication of periodicity,
but
not nearly as unambiguously as does the cepstrum. On the other
hand,
note that the rapid variations of the unvoiced spectra appear
random
with no periodic structure. This is typical of Fourier transforms
(periodograms)
of short segments of random signals. As a result, there is no
strong peak indicating periodicity as in the voiced case.
64 Homomorphic Speech Analysis
To illustrate the effect of liftering, the quefrencies above 5ms are
multiplied by zero and the quefrencies below 5 ms are multiplied
by 1
(with a short transition taper as shown in Figures 5.5(a) and
5.5(c)).
The DFT of the resulting modified cepstrum is plotted as the
smooth
curve that is superimposed on the shorttime spectra in Figures
5.5(b)
and 5.5(d), respectively. These slowly varying log spectra clearly
retain
the general spectral shape with peaks corresponding to the formant
resonance structure for the segment of speech under analysis.
Therefore,
a useful perspective is that by liftering the cepstrum, it is possible
to separate information about the vocal tract contributions from the
shorttime speech spectrum [88].
5.5 Application to Pitch Detection
The cepstrum was first applied in speech processing to determine
the
excitation parameters for the discretetime speech model of Figure
7.10.
Noll [83] applied the shorttime cepstrum to detect local
periodicity
(voiced speech) or the lack thereof (unvoiced speech). This is
illustrated
in Figure 5.6, which shows a plot that is very similar to the plot
first published by Noll [83]. On the left is a sequence of log short
time
spectra (rapidly varying curves) and on the right is the
corresponding
sequence of cepstra computed from the log spectra on the left.
The successive spectra and cepstra are for 50 ms segments
obtained by
moving the window in steps of 12.5 ms (100 samples at a sampling
rate
of 8000 samples/sec). From the discussion of Section 5.4, it is
apparent
that for the positions 1 through 5, the window includes only
unvoiced
speech, while for positions 6 and 7 the signal within the window is
partly voiced and partly unvoiced. For positions 8 through 15 the
window
only includes voiced speech. Note again that the rapid variations
of the unvoiced spectra appear random with no periodic structure.
On
the other hand, the spectra for voiced segments have a structure of
periodic ripples due to the harmonic structure of the quasiperiodic
segment of voiced speech. As can be seen from the plots on the
right,
the cepstrum peak at a quefrency of about 11–12 ms strongly
signals
voiced speech, and the quefrency of the peak is an accurate
estimate of
the pitch period during the corresponding speech interval.
5.5 Application to Pitch Detection 65
0 1 2 3 4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
0 5 10 15 20 25
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
ShortTime Log Spectra in Cepstrum Analysis
frequency in kHz
ShortTime Cepstra
time in ms
window number Fig. 5.6 Shorttime cepstra and corresponding
STFTs and homomorphicallysmoothed
spectra.
The essence of the pitch detection algorithm proposed by Noll is
to compute a sequence of shorttime cepstra and search each
successive
cepstrum for a peak in the quefrency region of the expected pitch
period. Presence of a strong peak implies voiced speech, and the
quefrency
location of the peak gives the estimate of the pitch period. As
in most modelbased signal processing applications of concepts
such
as the cepstrum, the pitch detection algorithm includes many
features
designed to handle cases that do not fit the underlying model
very well. For example, for frames 6 and 7 the cepstrum peak is
weak
66 Homomorphic Speech Analysis
corresponding to the transition from unvoiced to voiced speech. In
other
problematic cases, the peak at twice the pitch period may be
stronger
than the peak at the quefrency of the pitch period. Noll applied
temporal
continuity constraints to prevent such errors.
5.6 Applications to Pattern Recognition
Perhaps the most pervasive application of the cepstrum in speech
processing
is its use in pattern recognition systems such as design of vector
quantizers (VQ) and automatic speech recognizers (ASR). In such
applications, a speech signal is represented on a framebyframe
basis
by a sequence of shorttime cepstra. As we have shown, cepstra
can be
computed either by ztransform analysis or by DFT
implementation
of the characteristic system. In either case, we can assume that the
cepstrum vector corresponds to a gainnormalized (c[0] = 0)
minimumphase
vocal tract impulse response that is defined by the complex
cepstrum2
ˆh
[n] =
_
2c[n] 1≤ n ≤ nco
0 n < 0.
(5.18)
In problems such as VQ or ASR, a test pattern c[n] (vector of
cepstrum
values n = 1,2, . . . , nco) is compared against a similarly defined
reference pattern ￣c[n]. Such comparisons require a distance
measure.
For example, the Euclidean distance applied to the cepstrum would
give
D =
_nco
n=1
c[n] − ￣c[n]2 . (5.19a)
Equivalently in the frequency domain,
D =
1
2π
_ π
−π
___
log H(ej ˆω) − log ￣H (ej ˆω)
___
2
dˆω, (5.19b)
where log H(ej ˆω) is the log magnitude of the DTFT of h[n]
corresponding
to the complex cepstrum in (5.18) or the real part of the
2For minimumphase signals, the complex cepstrum satisfies h[n]
= 0 for n < 0. Since the
cepstrum is always the even part of the complex cepstrum, it
follows that ˆh[n] = 2c[n] for
n > 0.
5.6 Applications to Pattern Recognition 67
DTFT ofˆh[n] in (5.18). Thus, cepstrumbased comparisons are
strongly
related to comparisons of smoothed shorttime spectra. Therefore,
the
cepstrum offers an effective and flexible representation of speech
for
pattern recognition problems, and its interpretation as a difference
of
log spectra suggests a match to auditory perception mechanisms.
5.6.1 Compensation for Linear Filtering
Suppose that we have only a linearly filtered version of the speech
signal, y[n] = h[n] ∗ x[n], instead of x[n]. If the analysis window is
long
compared to the length of h[n], the shorttime cepstrum of one
frame
of the filtered speech signal y[n] will be3
c
(y)
ˆn [m] = c
(x)
ˆn [m] + c(h)[m], (5.20)
where c(h)[m] will appear essentially the same in each frame.
Therefore,
if we can estimate c(h)[n],4 which we assume is nontimevarying,
we can
obtain c
(x)
ˆn [m] at each frame from c
(y)
ˆn [m] by subtraction, i.e., c
(x)
n [m] =
c
(y)
ˆn [m] − c(h)[m]. This property is extremely attractive in situations
where the reference pattern ￣c[m] has been obtained under
different
recording or transmission conditions from those used to acquire the
test vector. In these circumstances, the test vector can be
compensated
for the effects of the linear filtering prior to computing the distance
measures used for comparison of patterns.
Another approach to removing the effects of linear distortions is to
observe that the cepstrum component due to the distortion is the
same
in each frame. Therefore, it can be removed by a simple first
difference
operation of the form:
Δc
(y)
ˆn [m] = c
(y)
ˆn [m] − c
(y)
ˆn−1[m]. (5.21)
It is clear that if c
(y)
ˆn [m] = c
(x)
ˆn [m] + c(h)[m] with c(h)[m] being independent
of ˆn, then Δc
(y)
ˆn [m] = Δc
(x)
ˆn [m], i.e., the linear distortion effects
are removed.
3 In this section, it will be useful to use somewhat more
complicated notation. Specifically,
we denote the cepstrum at analysis time ˆn of a signal x[n] as c
(x)
ˆn [m], where m denotes
the quefrency index of the cepstrum.
4 Stockham [124] showed how c(h)[m] for such linear distortions
can be estimated from the
signal y[n] by time averaging the log of the STFT.
68 Homomorphic Speech Analysis
Furui [39, 40] first noted that the sequence of cepstrum values has
temporal information that could be of value for a speaker
verification
system. He used polynomial fits to cepstrum sequences to extract
simple
representations of the temporal variation. The delta cepstrum as
defined in (5.21) is simply the slope of a firstorder polynomial fit
to
the cepstrum time evolution.
5.6.2 Liftered Cepstrum Distance Measures
In using linear predictive analysis (discussed in Chapter 6) to
obtain
cepstrum feature vectors for pattern recognition problems, it is
observed that there is significant statistical variability due to a
variety
of factors including shorttime analysis window position, bias
toward
harmonic peaks, and additive noise [63, 126]. A solution to this
problem
is to use weighted distance measures of the form:
D =
_nco
n=1
g2[n] c[n] − ￣c[n]2 , (5.22a)
which can be written as the Euclidean distance of liftered cepstra
D =
_nco
n=1
g[n]c[n] − g[n]￣c[n]2 . (5.22b)
Tohkura [126] found, for example, that when averaged over many
frames of speech and speakers, cepstrum values c[n] have zero
means
and variances on the order of 1/n2. This suggests that g[n] = n for
n = 1,2, . . . , nco could be used to equalize the contributions for
each
term to the cepstrum difference.
Juang et al. [63] observed that the variability due to the vagaries of
LPC analysis could be lessened by using a lifter of the form:
g[n]=1 + 0.5nco sin(πn/nco) n = 1,2, . . . , nco. (5.23)
Tests of weighted distance measures showed consistent
improvements
in automatic speech recognition tasks.
Itakura and Umezaki [55] used the group delay function to derive a
different cepstrum weighting function. Instead of g[n] = n for all n,
or
5.6 Applications to Pattern Recognition 69
the lifter of (5.23), Itakura proposed the lifter
g[n] = nse
−n2/2τ2
. (5.24)
This lifter has great flexibility. For example, if s = 0 we have
simply
lowpass liftering of the cepstrum. If s = 1 and τ is large, we have
essentially
g[n] = n for small n with high quefrency tapering. Itakura and
Umezaki [55] tested the group delay spectrum distance measure in
an
automatic speech recognition system. They found that for clean
test
utterances, the difference in recognition rate was small for different
values of s when τ ≈ 5 although performance suffered with
increasing
s for larger values of τ . This was attributed to the fact that for
larger s the group delay spectrum becomes very sharply peaked
and
thus more sensitive to small differences in formant locations.
However,
in test conditions with additive white noise and also with linear
filtering
distortions, recognition rates improved significantly with τ = 5 and
increasing values of the parameter s.
5.6.3 MelFrequency Cepstrum Coefficients
As we have seen, weighted cepstrum distance measures have a
directly
equivalent interpretation in terms of distance in the frequency
domain.
This is significant in light of models for human perception of
sound,
which, as noted in Chapter 3, are based upon a frequency analysis
performed in the inner ear. With this in mind, Davis and
Mermelstein
[27] formulated a new type of cepstrum representation that has
come to
be widely used and known as the melfrequency cepstrum
coefficients
(mfcc).
The basic idea is to compute a frequency analysis based upon a
filter bank with approximately critical band spacing of the filters
and
bandwidths. For 4 kHz bandwidth, approximately 20 filters are
used.
In most implementations, a shorttime Fourier analysis is done
first,
resulting in a DFT Xˆn[k] for analysis time ˆn. Then the DFT
values
are grouped together in critical bands and weighted by a triangular
weighting function as depicted in Figure 5.7. Note that the
bandwidths
in Figure 5.7 are constant for center frequencies below 1 kHz and
then
increase exponentially up to half the sampling rate of 4 kHz
resulting
70 Homomorphic Speech Analysis
0 1000 2000 3000 4000
0
0.005
0.01
frequency in Hz
Fig. 5.7 Weighting functions for Melfrequency filter bank.
in a total of 22 “filters.” The melfrequency spectrum at analysis
time
ˆn is defined for r = 1,2,...,R as
MFˆn[r] =
1
Ar
_Ur
k=Lr
Vr[k]Xˆn[k]2 , (5.25a)
where Vr[k] is the triangular weighting function for the rth filter
ranging
from DFT index Lr to Ur, where
Ar =
_Ur
k=Lr
Vr[k]2 (5.25b)
is a normalizing factor for the rth melfilter. This normalization is
built
into the weighting functions of Figure 5.7. It is needed so that a
perfectly
flat input Fourier spectrum will produce a flat melspectrum. For
each frame, a discrete cosine transform of the log of the magnitude
of
the filter outputs is computed to form the function mfccˆn[m], i.e.,
mfccˆn[m] =
1
R
_R
r=1
log(MFˆn[r])cos
_
2π
R
_
r +
1
2
_
m
_
. (5.26)
Typically, mfccˆn[m] is evaluated for a number of coefficients,
Nmfcc,
that is less than the number of melfilters, e.g., Nmfcc = 13 and R =
22.
Figure 5.8 shows the result of mfcc analysis of a frame of voiced
speech
in comparison with the shorttime Fourier spectrum, LPC spectrum
(discussed in Chapter 6), and a homomorphically smoothed
spectrum.
The large dots are the values of log(MFˆn[r]) and the line
interpolated
5.7 The Role of the Cepstrum 71
0 500 1000 1500 2000 2500 3000 3500 4000
−7
−6
−5
−4
−3
−2
−1
0
1
2
3
log magnitude
frequency in Hz
Shorttime Fourier Transform
Homomorphic smoothing,
LPC smoothing,
Mel cepstrum smoothing,
Fig. 5.8 Comparison of spectral smoothing methods.
between them is a spectrum reconstructed at the original DFT
frequencies.
Note that all these spectra are different, but they have in common
that they have peaks at the formant resonances. At higher
frequencies,
the reconstructed melspectrum of course has more smoothing due
to
the structure of the filter bank.
Note that the delta cepstrum idea expressed by (5.21) can be
applied
to mfcc to remove the effects of linear filtering as long as the
frequency
response of the distorting linear filter does not vary much across
each
of the melfrequency bands.
5.7 The Role of the Cepstrum
As discussed in this chapter, the cepstrum entered the realm of
speech
processing as a basis for pitch detection. It remains one of the most
effective indicators of voice pitch that have been devised. Because
the
72 Homomorphic Speech Analysis
vocal tract and excitation components are well separated in the
cepstrum,
it was natural to consider analysis techniques for estimation of
the vocal tract system as well [86, 88, 111]. While separation
techniques
based on the cepstrum can be very effective, the linear predictive
analysis
methods to be discussed in the next chapter have proven to be
more effective for a variety of reasons. Nevertheless, the cepstrum
concept
has demonstrated its value when applied to vocal tract system
estimates obtained by linear predictive analysis.
6
Linear Predictive Analysis
Linear predictive analysis is one of the most powerful and widely
used speech analysis techniques. The importance of this method
lies
both in its ability to provide accurate estimates of the speech
parameters
and in its relative speed of computation. In this chapter, we
present a formulation of the ideas behind linear prediction, and
discuss
some of the issues that are involved in using it in practical speech
applications.
6.1 Linear Prediction and the Speech Model
We come to the idea of linear prediction of speech by recalling the
source/system model that was introduced in Chapter 2, where the
sampled speech signal was modeled as the output of a linear,
slowly
timevarying system excited by either quasiperiodic impulses
(during
voiced speech), or random noise (during unvoiced speech). The
particular
form of the source/system model implied by linear predictive
analysis is depicted in Figure 6.1, where the speech model is the
part
inside the dashed box. Over short time intervals, the linear system
is
73
74 Linear Predictive Analysis
Fig. 6.1 Model for linear predictive analysis of speech signals.
described by an allpole system function of the form:
H(z) = S(z)
E(z)
= G
1 −
_p
k=1
akz
−k
. (6.1)
In linear predictive analysis, the excitation is defined implicitly by
the
vocal tract system model, i.e., the excitation is whatever is needed
to produce s[n] at the output of the system. The major advantage of
this model is that the gain parameter, G, and the filter coefficients
{ak} can be estimated in a very straightforward and
computationally
efficient manner by the method of linear predictive analysis.
For the system of Figure 6.1 with the vocal tract model of (6.1), the
speech samples s[n] are related to the excitation e[n] by the
difference
equation
s[n] =
_p
k=1
aks[n − k] + Ge[n]. (6.2)
A linear predictor with prediction coefficients, αk, is defined as a
system
whose output is
˜s[n] =
_p
k=1
αks[n − k], (6.3)
and the prediction error, defined as the amount by which ˜s[n] fails
to
exactly predict sample s[n], is
d[n] = s[n] − ˜s[n] = s[n] −
_p
k=1
αks[n − k]. (6.4)
6.1 Linear Prediction and the Speech Model 75
From (6.4) it follows that the prediction error sequence is the
output
of an FIR linear system whose system function is
A(z) = 1 −
_p
k=1
αkz
−k = D(z)
S(z) . (6.5)
It can be seen by comparing (6.2) and (6.4) that if the speech signal
obeys the model of (6.2) exactly, and if αk = ak, then d[n] = Ge[n].
Thus, the prediction error filter, A(z), will be an inverse filter for
the
system, H(z), of (6.1), i.e.,
H(z) = G
A(z) . (6.6)
The basic problem of linear prediction analysis is to determine the
set of predictor coefficients {αk} directly from the speech signal in
order to obtain a useful estimate of the timevarying vocal tract
system
through the use of (6.6). The basic approach is to find a set of
predictor
coefficients that will minimize the meansquared prediction error
over a short segment of the speech waveform. The resulting
parameters
are then assumed to be the parameters of the system function
H(z) in the model for production of the given segment of the
speech
waveform. This process is repeated periodically at a rate
appropriate
to track the phonetic variation of speech (i.e., order of 50–100
times
per second).
That this approach will lead to useful results may not be
immediately
obvious, but it can be justified in several ways. First, recall that
if αk = ak, then d[n] = Ge[n]. For voiced speech this means that
d[n]
would consist of a train of impulses, i.e., d[n] would be small
except at
isolated samples spaced by the current pitch period, P0. Thus,
finding
αks that minimize the meansquared prediction error seems
consistent
with this observation. A second motivation for this approach
follows
from the fact that if a signal is generated by (6.2) with nontime
varying
coefficients and excited either by a single impulse or by a
stationary
white noise input, then it can be shown that the predictor
coefficients
that result from minimizing the meansquared prediction error
(over
all time) are identical to the coefficients of (6.2). A third pragmatic
justification for using the minimum meansquared prediction error
as a
76 Linear Predictive Analysis
basis for estimating the model parameters is that this approach
leads to
an exceedingly useful and accurate representation of the speech
signal
that can be obtained by efficient solution of a set of linear
equations.
The shorttime average prediction error is defined as
Eˆn =
_
d2
ˆn[m]
_
=
__
sˆn[m] −
_p
k=1
αksˆn[m − k]
2_
, (6.7)
where sˆn[m] is a segment of speech that has been selected in a
neighborhood
of the analysis time ˆn, i.e.,
sˆn[m] = s[m + ˆn] − M1 ≤ m ≤ M2. (6.8)
That is, the time origin of the analysis segment is shifted to sample
ˆn of the entire signal. The notation _ _ denotes averaging over a
finite
number of samples. The details of specific definitions of the
averaging
operation will be discussed below.
We can find the values of αk that minimize Eˆn in (6.7) by setting
∂Eˆn/∂αi = 0, for i = 1,2, . . . , p, thereby obtaining the equations1
_p
k=1
˜αk _sˆn[m − i]sˆn[m − k]_ = _sˆn[m − i]sˆn[m]_ 1 ≤ i ≤ p, (6.9)
where the ˜αk are the values of αk that minimize Eˆn in (6.7).
(Since the
˜αk are unique, we will drop the tilde and use the notation αk to
denote
the values that minimize Eˆn.) If we define
ϕˆn[i,k] = _sˆn[m − i]sˆn[m − k]_ , (6.10)
then (6.9) can be written more compactly as2
_p
k=1
αkϕˆn[i,k] = ϕˆn[i, 0] i = 1,2, . . . , p. (6.11)
1 More accurately, the solutions of (6.7) only provide a stationary
point which can be shown
to be a minimum of Eˆn since Eˆn is a convex function of the αis.
2 The quantities ϕˆn[i,k] are in the form of a correlation function
for the speech segment
sˆn[m]. The details of the definition of the averaging operation
used in (6.10) have a
significant effect on the properties of the prediction coefficients
that are obtained by solving
(6.11).
6.2 Computing the Prediction Coefficients 77
If we know ϕˆn[i,k] for 1 ≤ i ≤ p and 0 ≤ k ≤ p, this set of p
equations
in p unknowns, which can be represented by the matrix equation:
Φα = ψ, (6.12)
can be solved for the vector α = {αk} of unknown predictor
coefficients
that minimize the average squared prediction error for the segment
sˆn[m].3 Using (6.7) and (6.9), the minimum meansquared
prediction
error can be shown to be [3, 5]
Eˆn = ϕˆn[0, 0] −
_p
k=1
αkϕˆn[0,k]. (6.13)
Thus, the total minimum meansquared error consists of a fixed
component
equal to the meansquared value of the signal segment minus a
term that depends on the predictor coefficients that satisfy (6.11),
i.e.,
the optimum coefficients reduce Eˆn in (6.13) the most.
To solve for the optimum predictor coefficients, we must first
compute
the quantities ϕˆn[i,k] for 1 ≤ i ≤ p and 0 ≤ k ≤ p. Once this
is done we only have to solve (6.11) to obtain the αks. Thus, in
principle, linear prediction analysis is very straightforward.
However,
the details of the computation of ϕˆn[i,k] and the subsequent
solution
of the equations are somewhat intricate and further discussion
is required.
6.2 Computing the Prediction Coefficients
So far we have not been explicit about the meaning of the
averaging
notation _ _ used to define the meansquared prediction error in
(6.7).
As we have stated, in a shorttime analysis procedure, the
averaging
must be over a finite interval. We shall see below that two methods
for
linear predictive analysis emerge out of a consideration of the
limits of
summation and the definition of the waveform segment sˆn[m].4
3 Although the αks are functions of ˆn (the time index at which
they are estimated) this
dependence will not be explicitly shown. We shall also find it
advantageous to drop the
subscripts ˆn on Eˆn, sˆn[m], and ϕˆn[i,k] when no confusion will
result.
4 These two methods, applied to the same speech signal yield
slightly different optimum
predictor coefficients.
78 Linear Predictive Analysis
6.2.1 The Covariance Method
One approach to computing the prediction coefficients is based on
the
definition
Eˆn =
_M2
m=−M1
(dˆn[m])2 =
_M2
m=−M1
_
sˆn[m] −
_p
k=1
αksˆn[m − k]
2
, (6.14)
where −M1 ≤ n ≤ M2. The quantities ϕˆn[i,k] needed in (6.11)
inherit
the same definition of the averaging operator, i.e.,
ϕˆn[i,k] =
_M2
m=−M1
sˆn[m − i]sˆn[m − k]
_
1 ≤ i ≤ p
0 ≤ k ≤ p.
(6.15)
Both (6.14) and (6.15) require values of sˆn[m] = s[m + ˆn] over
the
range −M1 − p ≤ m ≤ M2. By changes of index of summation,
(6.15)
can be expressed in the equivalent forms:
ϕˆn[i,k] =
M_2−i
m=−M1−i
sˆn[m]sˆn[m + i − k] (6.16a)
=
M_2−k
m=−M1−k
sˆn[m]sˆn[m + k − i], (6.16b)
from which it follows that ϕˆn[i,k] = ϕˆn[k, i].
Figure 6.2 shows the sequences that are involved in computing the
meansquared prediction error as defined by (6.14). The top part of
this figure shows a sampled speech signal s[m], and the box
denotes
a segment of that waveform selected around some time index ˆn.
The
second plot shows that segment extracted as a finitelength
sequence
sˆn[m] = s[m + ˆn] for −M1 − p ≤ m ≤ M2. Note the p “extra”
samples
(light shading) at the beginning that are needed to start the
prediction
error filter at time −M1. This method does not require any
assumption
about the signal outside the interval −M1 − p ≤ m ≤ M2 since
the samples sˆn[m] for −M1 − p ≤ m ≤ M2 are sufficient to
evaluate
ϕ[i,k] for all required values of i and k. The third plot shows the
prediction
error computed with the optimum predictor coefficients. Note
that in solving for the optimum prediction coefficients using
(6.11), the
6.2 Computing the Prediction Coefficients 79
Fig. 6.2 Illustration of windowing and prediction error for the
covariance method.
prediction error is implicitly computed over the range −M1 ≤ m ≤
M2
as required in (6.14) and (6.15). The minimum meansquared
prediction
error would be simply the sum of squares of all the samples shown.
In most cases of linear prediction, the prediction error sequence is
not
explicitly computed since solution of (6.11) does not require it.
The mathematical structure that defines the covariance method of
linear predictive analysis implies a number of useful properties of
the
solution. It is worthwhile to summarize them as follows:
(C.1) The meansquared prediction error satisfies Eˆn ≥ 0. With
this method, it is theoretically possible for the average
error to be exactly zero.
(C.2) The matrix Φ in (6.12) is a symmetric positivesemidefinite
matrix.
(C.3) The roots of the prediction error filter A(z) in (6.5) are
not guaranteed to lie within the unit circle of the zplane.
This implies that the vocal tract model filter (6.6) is not
guaranteed to be stable.
80 Linear Predictive Analysis
(C.4) As a result of (C.2), the equations (6.11) can be solved
efficiently using, for example, the well known Cholesky
decomposition of the covariance matrix Φ [3].
6.2.2 The Autocorrelation Method
Perhaps the most widely used method of linear predictive analysis
is called the autocorrelation method because the covariance
function
ϕˆn[i,k] needed in (6.11) reduces to the STACF φˆn[i − k] that
we discussed
in Chapter 4 [53, 54, 74, 78]. In the autocorrelation method, the
analysis segment sˆn[m] is defined as
sˆn[m] =
_
s[n + m]w[m] −M1 ≤ m ≤ M2
0 otherwise,
(6.17)
where the analysis window w[m] is used to taper the edges of the
segment
to zero. Since the analysis segment is defined by the windowing
of (6.17) to be zero outside the interval −M1 ≤ m ≤ M2, it follows
that
the prediction error sequence dˆn[m] can be nonzero only in the
range
−M1 ≤ m ≤ M2 + p. Therefore, Eˆn is defined as
Eˆn =
M_2+p
m=−M1
(dˆn[m])2 =
∞_
m=−∞
(dˆn[m])2. (6.18)
The windowing of (6.17) allows us to use the infinite limits to
signify
that the sum is over all nonzero values of dˆn[m]. Applying this
notion
to (6.16a) and (6.16b) leads to the conclusion that
ϕˆn[i,k] =
∞_
m=−∞
sˆn[m]sˆn[m + i − k] = φˆn[i − k]. (6.19)
Thus, ϕ[i,k] is a function only of i − k. Therefore, we can replace
ϕˆn[i,k] by φˆn[i − k], which is the STACF defined in Chapter 4
as
φˆn[k] =
∞_
m=−∞
sˆn[m]sˆn[m + k] = φˆn[−k]. (6.20)
The resulting set of equations for the optimum predictor
coefficients is
therefore
_p
k=1
αkφˆn[i − k] = φˆn[i] i = 1,2, . . . , p. (6.21)
6.2 Computing the Prediction Coefficients 81
Fig. 6.3 Illustration of windowing and prediction error for the
autocorrelation method.
Figure 6.3 shows the sequences that are involved in computing the
optimum prediction coefficients using the autocorrelation method.
The
upper plot shows the same sampled speech signal s[m] as in Figure
6.2
with a Hamming window centered at time index ˆn. The middle
plot
shows the result of multiplying the signal s[ˆn + m] by the window
w[m]
and redefining the time origin to obtain sˆn[m]. Note that the
zerovalued
samples outside the window are shown with light shading. The
third plot shows the prediction error computed using the optimum
coefficients. Note that for this segment, the prediction error (which
is
implicit in the solution of (6.21)) is nonzero over the range −M1 ≤
m ≤
M2 + p. Also note the lightly shaded p samples at the beginning.
These
samples can be large due to the fact that the predictor must predict
these samples from the zerovalued samples that precede the
windowed
segment sˆn[m]. It is easy to see that at least one of these first p
samples
of the prediction error must be nonzero. Similarly, the last p
samples of
the prediction error can be large due to the fact that the predictor
must
predict zerovalued samples from windowed speech samples. It
can also
be seen that at least one of these last p samples of the prediction
error
must be nonzero. For this reason, it follows that Eˆn, being the sum
of
82 Linear Predictive Analysis
squares of the prediction error samples, must always be strictly
greater
than zero.5
As in the case of the covariance method, the mathematical
structure
of the autocorrelation method implies a number of properties of the
solution, including the following:
(A.1) The meansquared prediction error satisfies Eˆn > 0. With
this method, it is theoretically impossible for the error
to be exactly zero because there will always be at least
one sample at the beginning and one at the end of the
prediction error sequence that will be nonzero.
(A.2) The matrix Φ in (6.12) is a symmetric positivedefinite
Toeplitz matrix [46].
(A.3) The roots of the prediction error filter A(z) in (6.5) are
guaranteed to lie within the unit circle of the zplane so
that the vocal tract model filter of (6.6) is guaranteed to
be stable.
(A.4) As a result of (A.2), the equations (6.11) can be solved
efficiently using the Levinson–Durbin algorithm, which,
because of its many implications, we discuss in more
detail in Section 6.3.
6.3 The Levinson–Durbin Recursion
As stated in (A.2) above, the matrix Φ in (6.12) is a symmetric
positivedefinite
Toeplitz matrix, which means that all the elements on a given
diagonal in the matrix are equal. Equation (6.22) shows the
detailed
structure of the matrix equation Φα = ψ for the autocorrelation
method.
¿
¿¿¿
φ[0] φ[1] · · · φ[p − 1]
φ[1] φ[0] · · · φ[p − 2]
· · · · · · · · · · · ·
φ[p − 1] φ[p − 2] · · · φ[0]
Q
QQQ\
¿
¿¿¿
α1
α2
· · ·
αp
Q
QQQ\
=
¿
¿¿¿
φ[1]
φ[2]
· · ·
φ[p]
Q
QQQ\
(6.22)
5For this reason, a tapering window is generally used in the
autocorrelation method.
6.3 The Levinson–Durbin Recursion 83
Note that the vector ψ is composed of almost the same
autocorrelation
values as comprise Φ. Because of the special structure of (6.22) it
is
possible to derive a recursive algorithm for inverting the matrix Φ.
That algorithm, known as the Levinson–Durbin algorithm, is
specified
by the following steps:
Levinson–Durbin Algorithm
E0 = φ[0] (D.1)
for i = 1,2, . . . , p
ki =
"
,φ[i] −
_i−1
j=1
α
(i−1)
j φ[i − j]
ü
ü
E(i−1) (D.2)
α
(i)
i = ki (D.3)
if i > 1 then for j = 1,2, . . . , i − 1
α
(i)
j = α
(i−1)
j
− kiα
(i−1)
i−j (D.4)
end
E(i) = (1 − k2
i )E(i−1) (D.5)
end
αj = α
(p)
j j = 1,2, . . . , p (D.6)
An important feature of the Levinson–Durbin algorithm is that it
determines by recursion the optimum ithorder predictor from the
optimum
(i − 1)thorder predictor, and as part of the process, all predictors
from order 0 (no prediction) to order p are computed along with
the corresponding meansquared prediction errors E(i).
Specifically, the
equations labeled (D.3) and (D.4) can be used to show that the
prediction
error system function satisfies
A(i)(z) = A(i−1)(z) − kiz
−iA(i−1)(z
−1). (6.23)
Defining the ithorder forward prediction error e(i)[n] as the output
of the prediction error filter with system function A(i)(z) and
b(i)[n] as the output of the ithorder backward prediction error
filter
B(i)(z) = z−iA(i)(z−1), (6.23) leads (after some manipulations) to
an
interpretation of the Levinson–Durbin algorithm in terms of a
lattice
filter structure as in Figure 6.4(a).
84 Linear Predictive Analysis
Fig. 6.4 Lattice structures derived from the Levinson–Durbin
recursion. (a) Prediction error
filter A(z). (b) Vocal tract filter H(z) = 1/A(z).
Also, by solving two equations in two unknowns recursively, it is
possible to start at the output of Figure 6.4(a) and work to the left
eventually computing s[n] in terms of e[n]. The lattice structure
corresponding
to this is shown in Figure 6.4(b). Lattice structures like this
can be derived from acoustic principles applied to a physical model
composed of concatenated lossless tubes [101]. If the input and
output
signals in such physical models are sampled at just the right
sampling
rate, the sampled signals are related by a transfer function identical
to
(6.6). In this case, the coefficients ki behave as reflection
coefficients at
the tube boundaries [3, 78, 101].
Note that the parameters ki for i = 1,2, . . . , p play a key role in
the Levinson–Durbin recursion and also in the lattice filter
interpretation.
Itakura and Saito [53, 54], showed that the parameters ki in the
Levinson–Durbin recursion and the lattice filter interpretation
obtained
from it also could be derived by looking at linear predictive
analysis
from a statistical perspective. They called the ki parameters,
PARCOR
(for partial correlation) coefficients [54], because they can be
computed
directly as a ratio of crosscorrelation values between the forward
and
6.4 LPC Spectrum 85
backward prediction errors at the output of the (i − 1)th stage of
prediction
in Figure 6.4(a), i.e.,
ki =
∞_
m=−∞
e(i−1)[m]b(i−1)[m − 1]
_ ∞_
m=−∞
!
e(i−1)[m]
"2
∞_
m=−∞
!
b(i−1)[m − 1]
"2
1/2 . (6.24)
In the PARCOR interpretation, each stage of Figure 6.4(a)
removes
part of the correlation in the input signal. The PARCOR
coefficients
computed using (6.24) are identical to the ki obtained as a result of
the
Levinson–Durbin algorithm. Indeed, Equation (D.2) in the
Levinson–
Durbin algorithm can be replaced by (6.24), and the result is an
algorithm
for transforming the PARCOR representation into the linear
predictor
coefficient representation.
The Levinson–Durbin formulation provides one more piece of
useful
information about the PARCOR coefficients. Specifically, from
Equation
(D.5) of the algorithm, it follows that, since E(i) = (1 − k2
i )E(i−1)
is strictly greater than zero for predictors of all orders, it must be
true
that −1 < ki < 1 for all i. It can be shown that this condition on the
PARCORs also guarantees that all the zeros of a prediction error
filter
A(i)(z) of any order must be strictly inside the unit circle of the z
plane
[74, 78].
6.4 LPC Spectrum
The frequencydomain interpretation of linear predictive analysis
provides
an informative link to our earlier discussions of the STFT and
cepstrum
analysis. The autocorrelation method is based on the shorttime
autocorrelation function, φˆn[m], which is the inverse discrete
Fourier
transform of the magnitudesquared of the STFT, Sˆn(ej ˆω)2, of
the
windowed speech signal sˆn[m] = s[n + m]w[m]. The values φˆn[m]
for
m = 0,1, . . . , p are used to compute the prediction coefficients and
gain,
which in turn define the vocal tract system function H(z) in (6.6).
Therefore, the magnitudesquared of the frequency response of this
system,
obtained by evaluating H(z) on the unit circle at angles 2πf/fs,
86 Linear Predictive Analysis
is of the form:
H(ej2πf/fs)2 =
__________
G
1 −
_p
k=1
αke
−j2πf/fs
__________
2
, (6.25)
and can be thought of as an alternative shorttime spectral
representation.
Figure 6.5 shows a comparison between shorttime Fourier
analysis
and linear predictive spectrum analysis for segments of voiced and
unvoiced speech. Figures 6.5(a) and 6.5(c) show the STACF, with
the
first 23 values plotted with a heavy line. These values are used to
compute the predictor coefficients and gain for an LPC model with
p = 22. The frequency responses of the corresponding vocal tract
system
models are computed using (6.25) where the sampling frequency is
fs = 16 kHz. These frequency responses are superimposed on the
corresponding
STFTs (shown in gray). As we have observed before, the rapid
variations with frequency in the STFT are due primarily to the
excitation,
while the overall shape is assumed to be determined by the
effects of glottal pulse, vocal tract transfer function, and radiation.
In
0 5 10 15
−0.5
0
0.5
1
0 5 10 15
−0.5
0
0.5
1
0 2000 4000 6000 8000
−100
−50
0
50
0 2000 4000 6000 8000
−100
−50
0
50
(a) Voiced Autocorrelation (b) Voiced LPC Spectrum
(c) Unvoiced Autocorrelation (d) Unvoiced LPC Spectrum
time in ms frequency in kHz
Fig. 6.5 Comparison of shorttime Fourier analysis with linear
predictive analysis.
6.4 LPC Spectrum 87
cepstrum analysis, the excitation effects are removed by lowpass
liftering
the cepstrum. In linear predictive spectrum analysis, the excitation
effects are removed by focusing on the lowtime autocorrelation
coefficients.
The amount of smoothing of the spectrum is controlled by
the choice of p. Figure 6.5 shows that a linear prediction model
with
p = 22 matches the general shape of the shorttime spectrum, but
does
not represent all its local peaks and valleys, and this is exactly
what is
desired.
The question naturally arises as to how p should be chosen.
Figure 6.6 offers a suggestion. In Figure 6.6(a) are shown the
STFT
(in gray) and the frequency responses of a 12thorder model (heavy
dark line) and a 40thorder model (thin dark line). Evidently, the
linear
predictive spectra tend to favor the peaks of the shorttime Fourier
transform. That this is true in general can be argued using the
Parseval
theorem of Fourier analysis [74]. This is in contrast to
homomorphic
Fig. 6.6 Effect of predictor order on: (a) estimating the spectral
envelope of a 4kHz bandwidth
signal; and (b) the normalized meansquared prediction error for
the same signal.
88 Linear Predictive Analysis
smoothing of the STFT, which tends toward an average of the
peaks
and valleys of the STFT. As an example, see Figure 5.8, which
shows
a comparison between the shorttime Fourier spectrum, the LPC
spectrum
with p = 12, a homomorphically smoothed spectrum using 13
cepstrum values, and a melfrequency spectrum.
Figure 6.6(b) shows the normalized meansquared prediction error
V (p) = E(p)/φ[0] as a function of p for the segment of speech used
to
produce Figure 6.6(a). Note the sharp decrease from p = 0 (no
prediction
implies V (0) = 1) to p = 1 and the less abrupt decrease thereafter.
Furthermore, notice that the meansquared error curve flattens
out above about p = 12 and then decreases modestly thereafter. The
choice p = 12 gives a good match to the general shape of the
STFT,
highlighting the formant structure imposed by the vocal tract filter
while ignoring the periodic pitch structure. Observe that p = 40
gives
a much different result. In this case, the vocal tract model is highly
influenced by the pitch harmonics in the shorttime spectrum. It
can
be shown that if p is increased to the length of the windowed
speech
segment, that H(ej2πf/fs)2 → Sˆn(ej2πf/fs)2, but as pointed out
in
(A.1) above, E(2M) does not go to zero [75].
In a particular application, the prediction order is generally fixed
at a value that captures the general spectral shape due to the glottal
pulse, vocal tract resonances, and radiation. From the acoustic
theory of speech production, it follows that the glottal pulse
spectrum
is lowpass, the radiation filtering is highpass, and the vocal tract
imposes a resonance structure that, for adult speakers, is comprised
of about one resonance per kilohertz of frequency [101]. For the
sampled
speech signal, the combination of the lowpass glottal pulse
spectrum
and the highpass filtering of radiation are usually adequately
represented by one or two additional complex pole pairs. When
coupled
with an estimate of one resonance per kilohertz, this leads to a
rule of thumb of p = 4 + fs/1000. For example, for a sampling rate
of
fs = 8000 Hz, it is common to use a predictor order of 10–12. Note
that
in Figure 6.5 where the sampling rate was fs = 16000 Hz, a
predictor
order of p = 22 gave a good representation of the overall shape and
resonance structure of the speech segment over the band from 0 to
8000 Hz.
6.5 Equivalent Representations 89
The example of Figure 6.6 illustrates an important point about
linear predictive analysis. In that example, the speaker was a
highpitched
female, and the wide spacing between harmonics causes the
peaks of the linear predictive vocal tract model to be biased toward
those harmonics. This effect is a serious limitation for highpitched
voices, but is less problematic for male voices where the spacing
between
harmonics is generally much smaller.
6.5 Equivalent Representations
The basic parameters obtained by linear predictive analysis are the
gain
G and the prediction coefficients {αk}. From these, a variety of
different
equivalent representations can be obtained. These different
representations
are important, particularly when they are used in speech coding,
because we often wish to quantize the model parameters for
efficiency
in storage or transmission of coded speech.
In this section, we give only a brief summary of the most important
equivalent representations.
6.5.1 Roots of Prediction Error System Function
Equation (6.5) shows that the system function of the prediction
error
filter is a polynomial in z−1 and therefore it can be represented in
terms
of its zeros as
A(z) = 1 −
_p
k=1
αkz
−k =
_p
k=1
(1 − zkz
−1). (6.26)
According to (6.6), the zeros of A(z) are the poles of H(z).
Therefore,
if the model order is chosen judiciously as discussed in the
previous
section, then it can be expected that roughly fs/1000 of the roots
will
be close in frequency (angle in the zplane) to the formant
frequencies.
Figure 6.7 shows an example of the roots (marked with ×) of a
12thorder
predictor. Note that eight (four complex conjugate pairs) of the
roots are close to the unit circle. These are the poles of H(z) that
model
the formant resonances. The remaining four roots lie well within
the
unit circle, which means that they only provide for the overall
spectral
shaping resulting from the glottal and radiation spectral shaping.
90 Linear Predictive Analysis
Fig. 6.7 Poles of H(z) (zeros of A(z)) marked with × and LSP roots
marked with ∗ and o.
The prediction coefficients are perhaps the most susceptible to
quantization
errors of all the equivalent representations. It is wellknown
that the roots of a polynomial are highly sensitive to errors in its
coefficients
— all the roots being a function of all the coefficients [89].
Since the pole locations are crucial to accurate representation of
the
spectrum, it is important to maintain high accuracy in the location
of
the zeros of A(z). One possibility, not often invoked, would be to
factor
the polynomial as in (6.26) and then quantize each root (magnitude
and angle) individually.
6.5.2 LSP Coefficients
A much more desirable alternative to quantization of the roots of
A(z)
was introduced by Itakura [52], who defined the line spectrum pair
(LSP) polynomials6
P(z) = A(z) + z
−(p+1)A(z
−1) (6.27a)
Q(z) = A(z) − z
−(p+1)A(z
−1). (6.27b)
6 Note the similarity to (6.23), from which it follows that P(z) and
Q(z) are system functions
of lattice filters obtained by extending Figure 6.4(a) with an
additional section with kp+1 =
∓1, respectively.
6.5 Equivalent Representations 91
This transformation of the linear prediction parameters is
invertible:
to recover A(z) from P(z) and Q(z) simply add the two equations to
obtain A(z) = (P(z) + Q(z))/2.
An illustration of the LSP representation is given in Figure 6.7,
which shows the roots of A(z), P(z), and Q(z) for a 12thorder
predictor.
This new representation has some very interesting and useful
properties that are confirmed by Figure 6.7 and summarized below
[52, 119].
(LSP.1) All the roots of P(z) and Q(z) are on the unit circle,
i.e., they can be represented as
P(z) =
p_+1
k=1
(1 − ejΩkz
−1) (6.28a)
Q(z) =
p_+1
k=1
(1 − ejΘkz
−1). (6.28b)
The normalized discretetime frequencies (angles in the
zplane), Ωk and Θk, are called the line spectrum frequencies
or LSFs. Knowing the p LSFs that lie in the
range 0 < ω < π is sufficient to completely define the
LSP polynomials and therefore A(z).
(LSP.2) If p is an even integer, then P(−1) = 0 and Q(1) = 0.
(LSP.3) The PARCOR coefficients corresponding to A(z) satisfy
ki < 1 if and only if the roots of P(z) and Q(z)
alternate on the unit circle, i.e., the LSFs are interlaced
over the range of angles 0 to π.
(LSP.4) The LSFs are close together when the roots of A(z)
are close to the unit circle.
These properties make it possible to represent the linear predictor
by quantized differences between the successive LSFs [119]. If
property
(LSP.3) is maintained through the quantization, the reconstructed
polynomial A(z) will still have its zeros inside the unit circle.
92 Linear Predictive Analysis
6.5.3 Cepstrum of Vocal Tract Impulse Response
One of the most useful alternative representations of the linear
predictor
is the cepstrum of the impulse response, h[n], of the vocal tract
filter. The impulse response can be computed recursively through
the
difference equation:
h[n] =
_p
k=1
αkh[n − k] + Gδ[n], (6.29)
or a closed form expression for the impulse response can be
obtained
by making a partial fraction expansion of the system function H(z).
As discussed in Section 5.3.3, since H(z) is a minimumphase
system
(all its poles inside the unit circle), it follows that the complex
cepstrum
of h[n] can be computed using (5.14), which would require that
the poles of H(z) be found by polynomial rooting. Then, any
desired
number of cepstrum values can be computed. However, because of
the
minimum phase condition, a more direct recursive computation is
possible
[101]. Equation (5.16) gives the recursion for computingˆh[n] from
the predictor coefficients and gain, and (5.17) gives the inverse
recursion
for computing the predictor coefficients from the complex
cepstrum of
the vocal tract impulse response. These relationships are
particularly
useful in speech recognition applications where a small number of
cepstrum
values is used as a feature vector.
6.5.4 PARCOR Coefficients
We have seen that the PARCOR coefficients are bounded by ±1.
This
makes them an attractive parameter for quantization.
We have mentioned that if we have a set of PARCOR coefficients,
we can simply use them at step (D.2) of the Levinson–Durbin
algorithm
thereby obtaining an algorithm for converting PARCORs to
predictor coefficients. By working backward through the
Levinson–
Durbin algorithm, we can compute the PARCOR coefficients from
a given set of predictor coefficients. The resulting algorithm is
given below.
6.5 Equivalent Representations 93
PredictortoPARCOR Algorithm
α
(p)
j = αj j = 1,2, . . . , p
kp = α
(p)
p (P.1)
for i = p,p − 1, . . . ,2
for j = 1,2, . . . , i − 1
α
(i−1)
j =
α
(i)
j + kiα
(i)
i−j
1 − k2
i
(P.2)
end
ki−1 = α
(i−1)
i−1 (P.3)
end
6.5.5 Log Area Coefficients
As mentioned above, the lattice filter representation of the vocal
tract
system is strongly related to concatenated acoustic tube models of
the
physics of sound propagation in the vocal tract. Such models are
characterized
by a set of tube crosssectional areas denoted Ai. These “area
function” parameters are useful alternative representations of the
vocal
tract model obtained from linear predictive analysis. Specifically,
the
log area ratio parameters are defined as
gi = log
_
Ai+1
Ai
_
= log
_
1 − ki
1 + ki
_
, (6.30)
where Ai and Ai+1 are the areas of two successive tubes and the
PARCOR
coefficient −ki is the reflection coefficient for sound waves
impinging
on the junction between the two tubes [3, 78, 101]. The inverse
transformation (from gi to ki) is
ki =
1 − egi
1 + egi
. (6.31)
The PARCOR coefficients can be converted to predictor
coefficients if
required using the technique discussed in Section 6.5.4.
Viswanathan and Makhoul [131] showed that the frequency
response
of a vocal tract filter represented by quantized log area ratio
coefficients
is relatively insensitive to quantization errors.
94 Linear Predictive Analysis
6.6 The Role of Linear Prediction
In this chapter, we have discussed many of the fundamentals of
linear
predictive analysis of speech. The value of linear predictive
analysis
stems from its ability to represent in a compact form the part of the
speech model associated with the vocal tract, which in turn is
closely
related to the phonemic representation of the speech signal. Over
40
years of research on linear predictive methods have yielded a
wealth of
knowledge that has been widely and effectively applied in almost
every
area of speech processing, but especially in speech coding, speech
synthesis,
and speech recognition. The present chapter and Chapters 3–5
provide the basis for the remaining chapters of this text, which
focus
on these particular application areas.
7
Digital Speech Coding
This chapter, the first of three on digital speech processing
applications,
focuses on specific techniques that are used in digital speech
coding.
We begin by describing the basic operation of sampling a speech
signal
and directly quantizing and encoding of the samples. The
remainder
of the chapter discusses a wide variety of techniques that represent
the
speech signal in terms of parametric models of speech production
and
perception.
7.1 Sampling and Quantization of Speech (PCM)
In any application of digital signal processing (DSP), the first step
is
sampling and quantization of the resulting samples into digital
form.
These operations, which comprise the process of AtoD
conversion,
are depicted for convenience in analysis and discussion in the
block
diagram of Figure 7.1.1
1 Current AtoD and DtoA converters use oversampling
techniques to implement the
functions depicted in Figure 7.1.
95
96 Digital Speech Coding
Sampler
Quantizer
Q{ }
xc (t) x[n] xˆ[n]
Encoder
c[n]
AtoD Converter
Decoder
c′[n] xˆ′[n]
Δ
T Δ
Reconstruction
Filter
DtoA Converter
) ( ˆ t x′c
Fig. 7.1 The operations of sampling and quantization.
The sampler produces a sequence of numbers x[n] = xc(nT), where
T is the sampling period and fs = 1/T is the sampling frequency.
Theoretically,
the samples x[n] are real numbers that are useful for theoretical
analysis, but never available for computations. The well known
Shannon sampling theorem states that a bandlimited signal can be
reconstructed exactly from samples taken at twice the highest
frequency
in the input signal spectrum (typically between 7 and 20 kHz for
speech
and audio signals) [89]. Often, we use a lowpass filter to remove
spectral
information above some frequency of interest (e.g., 4 kHz) and
then use
a sampling rate such as 8000 samples/s to avoid aliasing distortion
[89].
Figure 7.2 depicts a typical quantizer definition. The quantizer
simply
takes the real number inputs x[n] and assigns an output ˆx[n]
according
to the nonlinear discreteoutput mapping Q{ }.2 In the case of
the example of Figure 7.2, the output samples are mapped to one
of eight possible values, with samples within the peaktopeak
range
being rounded and samples outside the range being “clipped” to
either
the maximum positive or negative level. For samples within range,
the
quantization error, defined as
e[n] = ˆx[n] − x[n] (7.1)
2 Note that in this chapter, the notation ˆx[n] denotes the quantized
version of x[n], not the
complex cepstrum of x[n] as in Chapter 5.
7.1 Sampling and Quantization of Speech (PCM) 97
Fig. 7.2 8level midtread quantizer.
satisfies the condition
−Δ/2 < e[n] ≤ Δ/2, (7.2)
where Δ is the quantizer step size. A Bbit quantizer such as the
one
shown in Figure 7.2 has 2B levels (1 bit usually signals the sign).
Therefore, if the peaktopeak range is 2Xm, the step size will be
Δ = 2Xm/2B. Because the quantization levels are uniformly spaced
by
Δ, such quantizers are called uniform quantizers. Traditionally,
representation
of a speech signal by binarycoded quantized samples is called
pulsecode modulation (or just PCM) because binary numbers can
be
represented for transmission as on/off pulse amplitude modulation.
The block marked “encoder” in Figure 7.1 represents the assigning
of a binary code word to each quantization level. These code words
represent
the quantized signal amplitudes, and generally, as in Figure 7.2,
these code words are chosen to correspond to some convenient
binary
number system such that arithmetic can be done on the code words
as
98 Digital Speech Coding
if they were proportional to the signal samples.3 In cases where
accurate
amplitude calibration is desired, the step size could be applied as
depicted in the block labeled “Decoder” at the bottom of Figure
7.1.4
Generally, we do not need to be concerned with the fine distinction
between the quantized samples and the coded samples when we
have
a fixed step size, but this is not the case if the step size is adapted
from sampletosample. Note that the combination of the decoder
and lowpass reconstruction filter represents the operation of Dto
A
conversion [89].
The operation of quantization confronts us with a dilemma. Most
signals, speech especially, have a wide dynamic range, i.e., their
amplitudes
vary greatly between voiced and unvoiced sounds and between
speakers. This means that we need a large peaktopeak range so as
to
avoid clipping of the loudest sounds. However, for a given number
of
levels (bits), the step size Δ = 2Xm/2B is proportional to the peak
topeak
range. Therefore, as the peaktopeak range increases, (7.2) states
that the size of the maximum quantization error grows.
Furthermore,
for a uniform quantizer, the maximum size of the quantization
error
is the same whether the signal sample is large or small. For a given
peaktopeak range for a quantizer such as the one in Figure 7.2 we
can decrease the quantization error only by adding more levels
(bits).
This is the fundamental problem of quantization.
The data rate (measured in bits/second) of sampled and quantized
speech signals is I = B · fs. The standard values for sampling and
quantizing
sound signals (speech, singing, instrumental music) are B = 16
and fs = 44.1 or 48 kHz. This leads to a bitrate of I = 16 · 44100 =
705,600 bits/s.5 This value is more than adequate and much more
than
desired for most speech communication applications. The bit rate
can
be lowered by using fewer bits/sample and/or using a lower
sampling
rate; however, both these simple approaches degrade the perceived
quality of the speech signal. This chapter deals with a wide range
of
3 In a real AtoD converter, the sampler, quantizer, and coder are
all integrated into one
system, and only the binary code words c[n] are available.
4 The _ denotes the possibility of errors in the codewords. Symbol
errors would cause additional
error in the reconstructed samples.
5 Even higher rates (24 bits and 96 kHz sampling rate) are used for
highquality audio.
7.1 Sampling and Quantization of Speech (PCM) 99
techniques for significantly reducing the bit rate while maintaining
an
adequate level of speech quality.
7.1.1 Uniform Quantization Noise Analysis
A more quantitative description of the effect of quantization can be
obtained using random signal analysis applied to the quantization
error.
If the number of bits in the quantizer is reasonably high and no
clipping
occurs, the quantization error sequence, although it is completely
determined by the signal amplitudes, nevertheless behaves as if it
is a
random signal with the following properties [12]:
(Q.1) The noise samples appear to be6 uncorrelated with the
signal samples.
(Q.2) Under certain assumptions, notably smooth input probability
density functions and high rate quantizers, the
noise samples appear to be uncorrelated from sampletosample,
i.e., e[n] acts like a white noise sequence.
(Q.3) The amplitudes of noise samples are uniformly distributed
across the range −Δ/2 < e[n] ≤ Δ/2, resulting
in average power σ2
e =Δ2/12.
These simplifying assumptions allow a linear analysis that yields
accurate
results if the signal is not too coarsely quantized. Under these
conditions,
it can be shown that if the output levels of the quantizer are
optimized, then the quantizer error will be uncorrelated with the
quantizer
output (however not the quantizer input, as commonly stated).
These results can be easily shown to hold in the simple case of a
uniformly
distributed memoryless input and Bennett has shown how the
result can be extended to inputs with smooth densities if the bit rate
is assumed high [12].
With these assumptions, it is possible to derive the following
formula for the signaltoquantizingnoise ratio (in dB) of a Bbit
6 By “appear to be” we mean that measured correlations are small.
Bennett has shown that
the correlations are only small because the error is small and, under
suitable conditions,
the correlations are equal to the negative of the error variance [12].
The condition “uncorrelated”
often implies independence, but in this case the error is a
deterministic function
of the input and hence it cannot be independent of the input.
100 Digital Speech Coding
uniform quantizer:
SNRQ = 10log
_
σ2x
σ2
e
_
= 6.02B + 4.78 − 20log10
_
Xm
σx
_
, (7.3)
where σx and σe are the rms values of the input signal and
quantization
noise samples, respectively. The formula of Equation (7.3) is an
increasingly
good approximation as the bit rate (or the number of quantization
levels) gets large. It can be way off, however, if the bit rate is not
large
[12]. Figure 7.3 shows a comparison of (7.3) with signalto
quantizationnoise
ratios measured for speech signals. The measurements were done
by quantizing 16bit samples to 8, 9, and 10 bits. The faint dashed
lines are from (7.3) and the dark dashed lines are measured values
for
uniform quantization. There is good agreement between these
graphs
indicating that (7.3) is a reasonable estimate of SNR.
Note that Xm is a fixed parameter of the quantizer, while σx
depends on the input signal level. As signal level increases, the
ratio
Xm/σx decreases moving to the left in Figure 7.3. When σx gets
close
to Xm, many samples are clipped, and the assumptions underlying
(7.3) no longer hold. This accounts for the precipitous fall in SNR
for
1<Xm/σx < 8.
0
10
20
30
40
50
60
8
9
10
Comparison of Law and Uniform Quantizers
SNR in dB
Fig. 7.3 Comparison of μlaw and linear quantization for B = 8,9,
10: Equation (7.3)—light
dashed lines. Measured uniform quantization SNR — dark dashed
lines. μlaw (μ = 100)
compression — solid lines.
7.1 Sampling and Quantization of Speech (PCM) 101
Also, it should be noted that (7.3) and Figure 7.3 show that with
all other parameters being fixed, increasing B by 1 bit (doubling
the
number of quantization levels) increases the SNR by 6 dB. On the
other
hand, it is also important to note that halving σx decreases the SNR
by 6 dB. In other words, cutting the signal level in half is like
throwing
away one bit (half of the levels) of the quantizer. Thus, it is
exceedingly
important to keep input signal levels as high as possible without
clipping.
7.1.2 μLaw Quantization
Also note that the SNR curves in Figure 7.3 decrease linearly with
increasing values of log10(Xm/σx). This is because the size of the
quantization
noise remains the same as the signal level decreases. If the
quantization error were proportional to the signal amplitude, the
SNR
would be constant regardless of the size of the signal. This could
be
achieved if log(x[n]) were quantized instead of x[n], but this is not
possible
because the log function is illbehaved for small values of x[n]. As
a compromise, μlaw compression (of the dynamic range) of the
signal,
defined as
y[n] = Xm
log
!
1 + μ
x[n]
Xm
"
log(1 + μ)
· sign(x[n]) (7.4)
can be used prior to uniform quantization [118]. Quantization of μ
law
compressed signals is often termed logPCM. If μ is large (> 100)
this
nonlinear transformation has the effect of distributing the effective
quantization levels uniformly for small samples and
logarithmically over
the remaining range of the input samples. The flat curves in Figure
7.3
show that the signaltonoise ratio with μlaw compression remains
relatively
constant over a wide range of input levels.
μlaw quantization for speech is standardized in the CCITT
G.711 standard. In particular, 8bit, μ = 255, logPCM with fs =
8000 samples/s is widely used for digital telephony. This is an
example
of where speech quality is deliberately compromised in order to
achieve a much lower bit rate. Once the bandwidth restriction has
been
imposed, 8bit logPCM introduces little or no perceptible
distortion.
102 Digital Speech Coding
This configuration is often referred to as “toll quality” because
when
it was introduced at the beginning of the digital telephony era, it
was
perceived to render speech equivalent to speech transmitted over
the
best long distance lines. Nowadays, readily available hardware
devices
convert analog signals directly into binarycoded μlaw samples
and
also expand compressed samples back to uniform scale for
conversion
back to analog signals. Such an AtoD/DtoA device is called
a encoder/decoder or “codec.” If the digital output of a codec is
used
as input to a speech processing algorithm, it is generally necessary
to
restore the linearity of the amplitudes through the inverse of (7.4).
7.1.3 NonUniform and Adaptive Quantization
μlaw compression is an example of nonuniform quantization. It is
based on the intuitive notion of constant percentage error. A more
rigorous
approach is to design a nonuniform quantizer that minimizes the
meansquared quantization error. To do this analytically, it is
necessary
to know the probability distribution of the signal sample values so
that
the most probable samples, which for speech are the low amplitude
samples, will incur less error than the least probable samples. To
apply
this idea to the design of a nonuniform quantizer for speech
requires
an assumption of an analytical form for the probability distribution
or
some algorithmic approach based on measured distributions. The
fundamentals
of optimum quantization were established by Lloyd [71] and
Max [79]. Paez and Glisson [91] gave an algorithm for designing
optimum
quantizers for assumed Laplace and Gamma probability densities,
which are useful approximations to measured distributions for
speech.
Lloyd [71] gave an algorithm for designing optimum nonuniform
quantizers
based on sampled speech signals.7 Optimum nonuniform
quantizers
can improve the signaltonoise ratio by as much as 6 dB over
μlaw quantizers with the same number of bits, but with little or no
improvement in perceived quality of reproduction, however.
7 Lloyd’s work was initially published in a Bell Laboratories
Technical Note with portions
of the material having been presented at the Institute of
Mathematical Statistics Meeting
in Atlantic City, New Jersey in September 1957. Subsequently this
pioneering work was
published in the open literature in March 1982 [71].
7.2 Digital Speech Coding 103
Another way to deal with the wide dynamic range of speech is to
let
the quantizer step size vary with time. When the quantizer in a
PCM
system is adaptive, the system is called an adaptive PCM (APCM)
system. For example, the step size could satisfy
Δ[n] = Δ0(Eˆn)1/2, (7.5)
where Δ0 is a constant and Eˆn is the shorttime energy of the
speech
signal defined, for example, as in (4.6). Such adaption of the step
size
is equivalent to fixed quantization of the signal after division by
the
rms signal amplitude (Eˆn)1/2. With this definition, the step size
will
go up and down with the local rms amplitude of the speech signal.
The adaptation speed can be adjusted by varying the analysis
window
size. If the step size control is based on the unquantized signal
samples, the quantizer is called a feedforward adaptive quantizer,
and
the step size information must be transmitted or stored along with
the
quantized sample, thereby adding to the information rate. To
amortize
this overhead, feedforward quantization is usually applied to
blocks of
speech samples. On the other hand, if the step size control is based
on
past quantized samples, the quantizer is a feedback adaptive
quantizer.
Jayant [57] studied a class of feedback adaptive quantizers where
the
step size Δ[n] is a function of the step size for the previous sample,
Δ[n − 1]. In this approach, if the previous sample is quantized
using
step size Δ[n − 1] to one of the low quantization levels, then the
step
size is decreased for the next sample. On the other hand, the step
size
is increased if the previous sample was quantized in one of the
highest
levels. By basing the adaptation on the previous sample, it is not
necessary
to transmit the step size. It can be derived at the decoder by
the same algorithm as used for encoding.
Adaptive quantizers can achieve about 6 dB improvement
(equivalent
to adding one bit) over fixed quantizers [57, 58].
7.2 Digital Speech Coding
As we have seen, the data rate of sampled and quantized speech is
B · fs. Virtually perfect perceptual quality is achievable with high
data
104 Digital Speech Coding
rate.8 By reducing B or fs, the bit rate can be reduced, but
perceptual
quality may suffer. Reducing fs requires bandwidth reduction,
and reducing B too much will introduce audible distortion that may
resemble random noise. The main objective in digital speech
coding
(sometimes called speech compression) is to lower the bit rate
while
maintaining an adequate level of perceptual fidelity. In addition to
the
two dimensions of quality and bit rate, the complexity (in terms of
digital computation) is often of equal concern.
Since the 1930s, engineers and speech scientists have worked
toward
the ultimate goal of achieving more efficient representations. This
work
has led to a basic approach in speech coding that is based on the
use ofDSP
techniques and the incorporation of knowledge of speech
production and
perception into the quantization process. There are many ways that
this
can be done. Traditionally, speech coding methods have been
classified
according to whether they attempt to preserve the waveform of the
speech signal or whether they only seek to maintain an acceptable
level
of perceptual quality and intelligibility. The former are generally
called
waveform coders, and the latter modelbased coders. Modelbased
coders
are designed for obtaining efficient digital representations of
speech and
only speech. Straightforward sampling and uniform quantization
(PCM)
is perhaps the only pure waveform coder. Nonuniform
quantization
and adaptive quantization are simple attempts to build in properties
of the speech signal (timevarying amplitude distribution) and
speech
perception (quantization noise is masked by loud sounds), but
these
extensions still are aimed at preserving the waveform.
Most modern modelbased coding methods are based on the notion
that the speech signal can be represented in terms of an excitation
signal
and a timevarying vocal tract system. These two components are
quantized, and then a “quantized” speech signal can be
reconstructed
by exciting the quantized filter with the quantized excitation.
Figure 7.4
illustrates why linear predictive analysis can be fruitfully applied
for
modelbased coding. The upper plot is a segment of a speech
signal.
8 If the signal that is reconstructed from the digital representation
is not perceptibly different
from the original analog speech signal, the digital representation is
often referred to as
being “transparent.”
7.2 Digital Speech Coding 105
0 100 200 300 400
−1
0
1
0 100 200 300 400
−0.2
0
0.2
Speech Waveform
Prediction Error Sequence
sample index
Fig. 7.4 Speech signal and corresponding prediction error signal
for a 12thorder predictor.
The lower plot is the output of a linear prediction error filter A(z)
(with p = 12) that was derived from the given segment by the
techniques
of Chapter 6. Note that the prediction error sequence amplitude
of Figure 7.4 is a factor of five lower than that of the signal itself,
which
means that for a fixed number of bits, this segment of the
prediction
error could be quantized with a smaller step size than the
waveform
itself. If the quantized prediction error (residual) segment is used
as
input to the corresponding vocal tract filter H(z) = 1/A(z), an
approximation
to the original segment of speech would be obtained.9 While
direct implementation of this idea does not lead to a practical
method
of speech coding, it is nevertheless suggestive of more practical
schemes
which we will discuss.
Today, the line between waveform coders and modelbased coders
is not distinct. A more useful classification of speech coders
focusses
on how the speech production and perception models are
incorporated
into the quantization process. One class of systems simply attempts
to
extract an excitation signal and a vocal tract system from the
speech
signal without any attempt to preserve a relationship between the
waveforms of the original and the quantized speech. Such systems
are
attractive because they can be implemented with modest
computation.
9To implement this timevarying inverse filtering/reconstruction
requires care in fitting the
segments of the signals together at the block boundaries. Overlap
add methods [89] can
be used effectively for this purpose.
106 Digital Speech Coding
These coders, which we designate as openloop coders, are also
called
vocoders (voice coder) since they are based on the principles
established
by H. Dudley early in the history of speech processing research
[30]. A second class of coders employs the source/system model
for
speech production inside a feedback loop, and thus are called
closedloop
coders. These compare the quantized output to the original input
and attempt to minimize the difference between the two in some
prescribed
sense. Differential PCM systems are simple examples of this
class, but increased availability of computational power has made
it
possible to implement much more sophisticated closedloop
systems
called analysisbysynthesis coders. Closed loop systems, since
they
explicitly attempt to minimize a timedomain distortion measure,
often
do a good job of preserving the speech waveform while employing
many
of the same techniques used in openloop systems.
7.3 ClosedLoop Coders
7.3.1 Predictive Coding
The essential features of predictive coding of speech were set forth
in a
classic paper by Atal and Schroeder [5], although the basic
principles of
predictive quantization were introduced by Cutler [26]. Figure 7.5
shows
a general block diagram of a large class of speech coding systems
that
are called adaptive differential PCM (ADPCM) systems. These
systems
are generally classified as waveform coders, but we prefer to
emphasize
that they are closedloop, modelbased systems that also preserve
the
waveform. The reader should ignore initially the blocks concerned
with
adaptation and all the dotted control lines and focus instead on the
core
DPCM system, which is comprised of a feedback structure that
includes
the blocks labeled Q and P. The quantizer Q can have a number of
levels ranging from 2 to much higher, it can be uniform or non
uniform,
and it can be adaptive or not, but irrespective of the type of
quantizer,
the quantized output can be expressed as ˆ d[n] = d[n] + e[n],
where e[n]
is the quantization error. The block labeled P is a linear predictor,
so
˜x[n] =
_p
k=1
αk ˆx[n − k], (7.6)
7.3 ClosedLoop Coders 107
Fig. 7.5 General block diagram for adaptive differential PCM
(ADPCM).
i.e., the signal ˜x[n] is predicted based on p past samples of the
signal
ˆx[n].10 The signal d[n] = x[n] − ˜x[n], the difference between the
input
and the predicted signal, is the input to the quantizer. Finally, the
10 In a feedforward adaptive predictor, the predictor parameters
are estimated from the
input signal x[n], although, as shown in Figure 7.5, they can also
be estimated from past
samples of the reconstructed signal ˆx[n]. The latter would be
termed a feedback adaptive
predictor.
108 Digital Speech Coding
relationship between ˆx[n] and ˆ d[n] is
ˆx[n] =
_p
k=1
αk ˆx[n − k] + ˆ d[n], (7.7)
i.e., the signal ˆx[n] is the output of what we have called the “vocal
tract filter,” and ˆ d[n] is the quantized excitation signal. Finally,
since
ˆx[n] = ˜x[n] + ˆ d[n] it follows that
ˆx[n] = x[n] + e[n]. (7.8)
This is the key result for DPCM systems. It says that no matter
what
predictor is used in Figure 7.5, the quantization error in ˆx[n] is
identical
to the quantization error in the quantized excitation signal ˆ d[n].
If the prediction is good, then the variance of d[n] will be less than
the variance of x[n] so it will be possible to use a smaller step size
and
therefore to reconstruct ˆx[n] with less error than if x[n] were
quantized
directly. The feedback structure of Figure 7.5 contains within it a
source/system speech model, which, as shown in Figure 7.5(b), is
the
system needed to reconstruct the quantized speech signal ˆx[n]
from the
coded difference signal input.
From (7.8), the signaltonoise ratio of the ADPCM system is
SNR = σ2x
/σ2
e . A simple, but informative modification leads to
SNR = σ2x
σ2
d
· σ2
d
σ2
e
= GP · SNRQ, (7.9)
where the obvious definitions apply.
SNRQ is the signaltonoise ratio of the quantizer, which, for fine
quantization (i.e., large number of quantization levels with small
step
size), is given by (7.3). The number of bits B determines the bit
rate
of the coded difference signal. As shown in Figure 7.5, the
quantizer
can be either feedforward or feedbackadaptive. As indicated, if
the
quantizer is feedforward adaptive, step size information must be
part
of the digital representation.
The quantity GP is called the prediction gain, i.e., if GP > 1 it
represents an improvement gained by placing a predictor based on
the
speech model inside the feedback loop around the quantizer.
Neglecting
7.3 ClosedLoop Coders 109
the correlation between the signal and the quantization error, it can
be
shown that
GP = σ2x
σ2
d
=
1
1 −
_p
k=1
αkρ[k]
, (7.10)
where ρ[k] = φ[k]/φ[0] is the normalized autocorrelation function
used
to compute the optimum predictor coefficients. The predictor can
be
either fixed or adaptive, and adaptive predictors can be either
feedforward
or feedback (derived by analyzing past samples of the quantized
speech signal reconstructed as part of the coder). Fixed predictors
can be designed based on longterm average correlation functions.
The
lower figure in Figure 7.6 shows an estimated longterm
autocorrelation
function for speech sampled at an 8 kHz sampling rate. The upper
plot shows 10log10GP computed from (7.10) as a function of
predictor
order p. Note that a firstorder fixed predictor can yield about 6 dB
prediction gain so that either the quantizer can have 1 bit less or
the
0 2 4 6 8 10 12
0
5
10
G
p
in dB
0 2 4 6 8 10 12
−0.5
0
0.5
1
predictor order p
autocorrelation lag
Fig. 7.6 Longterm autocorrelation function ρ[m] (lower plot) and
corresponding prediction
gain GP .
110 Digital Speech Coding
reconstructed signal can have a 6 dB higher SNR. Higherorder
fixed
predictors can achieve about 4 dB more prediction gain, and
adapting
the predictor at a phoneme rate can produce an additional 4 dB of
gain
[84]. Great flexibility is inherent in the block diagram of Figure
7.5, and
not surprisingly, many systems based upon the basic principle of
differential
quantization have been proposed, studied, and implemented
as standard systems. Here we can only mention a few of the most
important.
7.3.2 Delta Modulation
Delta modulation (DM) systems are the simplest differential
coding
systems since they use a 1bit quantizer and usually only a first
order
predictor. DM systems originated in the classic work of Cutler [26]
and deJager [28]. While DM systems evolved somewhat
independently
of the more general differential coding methods, they nevertheless
fit
neatly into the general theory of predictive coding. To see how
such
systems can work, note that the optimum predictor coefficient for a
firstorder predictor is α1 = ρ[1], so from (7.10) it follows that the
prediction
gain for a firstorder predictor is
GP =
1
1 − ρ2[1] . (7.11)
Furthermore, note the nature of the longterm autocorrelation
function
in Figure 7.6. This correlation function is for fs = 8000 samples/s.
If
the same bandwidth speech signal were oversampled at a
sampling
rate higher than 8000 samples/s, we could expect that the
correlation
value ρ[1] would lie on a smooth curve interpolating between
samples
0 and 1 in Figure 7.6, i.e., as fs gets large, both ρ[1] and α1
approach
unity and GP gets large. Thus, even though a 1bit quantizer has a
very low SNR, this can be compensated by the prediction gain due
to
oversampling.
Delta modulation systems can be implemented with very simple
hardware, and in contrast to more general predictive coding
systems,
they are not generally implemented with block processing. Instead,
adaptation algorithms are usually based on the bit stream at the
output
of the 1bit quantizer. The simplest delta modulators use a very
high
7.3 ClosedLoop Coders 111
sampling rate with a fixed predictor. Their bit rate, being equal to
fs,
must be very high to achieve good quality reproduction. This
limitation
can be mitigated with only modest increase in complexity by using
an
adaptive 1bit quantizer [47, 56]. For example, Jayant [56] showed
that
for bandlimited (3 kHz) speech, adaptive delta modulation (ADM)
with
bit rate (sampling rate) 40 kbits/s has about the same SNR as 6bit
log PCM sampled at 6.6 kHz. Furthermore, he showed that
doubling
the sampling rate of his ADM system improved the SNR by about
10 dB.
Adaptive delta modulation has the following advantages: it is very
simple to implement, only bit synchronization is required in
transmission,
it has very low coding delay, and it can be adjusted to be robust to
channel errors. For these reasons, adaptive delta modulation
(ADM) is
still used for terminalcostsensitive digital transmission
applications.
7.3.3 Adaptive Differential PCM Systems
Differential quantization with a multibit quantizer is called
differential
PCM (DPCM). As depicted in Figure 7.5, adaptive differential
PCM systems can have any combination of adaptive or fixed
quantizers
and/or predictors. Generally, the operations depicted in Figure 7.5
are implemented on short blocks of input signal samples, which
introduces
delay. A careful comparison of a variety of ADPCM systems by
Noll [84] gives a valuable perspective on the relative contributions
of
the quantizer and predictor. Noll compared log PCM, adaptive
PCM,
fixed DPCM, and three ADPCM systems all using quantizers of 2,
3,
4, and 5bits with 8 kHz sampling rate. For all bit rates (16, 24, 32,
and
40 kbps), the results were as follows11:
(1) Log PCM had the lowest SNR.
(2) Adapting the quantizer (APCM) improved SNR by 6 dB.
(3) Adding firstorder fixed or adaptive prediction improved the
SNR by about 4 dB over APCM.
11 The bit rates mentioned do not include overhead information for
the quantized prediction
coefficients and quantizer step sizes. See Section 7.3.3.1 for a
discussion of this issue.
112 Digital Speech Coding
(4) A fourthorder adaptive predictor added about 4 dB, and
increasing the predictor order to 12 added only 2 dB more.
The superior performance of ADPCM relative to logPCM and its
relatively low computational demands have led to several standard
versions
(ITU G.721, G.726, G.727), which are often operated at 32 kbps
where quality is superior to logPCM (ITU G.711) at 64 kbps.
The classic paper by Atal and Schroeder [5] contained a number
of ideas that have since been applied with great success in adaptive
predictive coding of speech. One of these concerns the
quantization
noise introduced by ADPCM, which we have seen is simply added
to
the input signal in the reconstruction process. If we invoke the
white
noise approximation for quantization error, it is clear that the noise
will
be most prominent in the speech at high frequencies where speech
spectrum
amplitudes are low. A simple solution to this problem is to
preemphasize
the speech signal before coding, i.e., filter the speech with a
linear filter that boosts the highfrequency (HF) part of the
spectrum.
After reconstruction of the preemphasized speech, it can be
filtered
with the inverse system to restore the spectral balance, and in the
process, the noise spectrum will take on the shape of the de
emphasis
filter spectrum. A simple preemphasis system of this type has
system
function (1 − γz−1) where γ is less than one.12 A more
sophisticated
application of this idea is to replace the simple fixed preemphasis
filter
with a timevarying filter designed to shape the quantization noise
spectrum. An equivalent approach is to define a new feedback
coding
system where the quantization noise is computed as the difference
between the input and output of the quantizer and then shaped
before
adding to the prediction residual. All these approaches raise the
noise
spectrum at low frequencies and lower it at high frequencies, thus
taking
advantage of the masking effects of the prominent low frequencies
in speech [2].
Another idea that has been applied effectively in ADPCM systems
as well as analysisbysynthesis systems is to include a longdelay
predictor
to capture the periodicity as well as the shortdelay correlation
12Values around γ = 0.4 for fs = 8 kHz work best. If γ is too close
to one, the low frequency
noise is emphasized too much.
7.3 ClosedLoop Coders 113
inherent in voiced speech. Including the simplest longdelay
predictor
would change (7.6) to
˜x[n] = βˆx[n − M] +
_p
k=1
αk(ˆx[n] − βˆx[n − M]). (7.12)
The parameter M is essentially the pitch period (in samples) of
voiced
speech, and the parameter β accounts for amplitude variations
between
periods. The predictor parameters in (7.12) cannot be jointly
optimized.
One suboptimal approach is to estimate β and M first by
determining M that maximizes the correlation around the expected
pitch period and setting β equal to the normalized correlation at M.
Then the shortdelay predictor parameters αk are estimated from
the
output of the longdelay prediction error filter.13 When this type of
predictor is used, the “vocal tract system” used in the decoder
would
have system function
H(z) =
"
"""",
1
1 −
_p
k=1
αkz
−k
ü
¸¸¸¸ü
_
1
1 − βz−M
_
. (7.13)
Adding longdelay prediction takes more information out of the
speech
signal and encodes it in the predictor. Even more complicated
predictors
with multiple delays and gains have been found to improve the
performance of ADPCM systems significantly [2].
7.3.3.1 Coding the ADPCM Parameters
A glance at Figure 7.5 shows that if the quantizer and predictor are
fixed or feedbackadaptive, then only the coded difference signal is
needed to reconstruct the speech signal (assuming that the decoder
has the same coefficients and feedbackadaptation algorithms).
Otherwise,
the predictor coefficients and quantizer step size must be included
as auxiliary data (side information) along with the quantized
difference
13 Normally, this analysis would be performed on the input speech
signal with the resulting
predictor coefficients applied to the reconstructed signal as in
(7.12).
114 Digital Speech Coding
signal. The difference signal will have the same sampling rate as
the
input signal. The step size and predictor parameters will be
estimated
and changed at a much lower sampling rate, e.g., 50–100 times/s.
The
total bit rate will be the sum of all the bit rates.
Predictor Quantization The predictor parameters must be
quantized
for efficient digital representation of the speech signal. As
discussed
in Chapter 6, the shortdelay predictors can be represented in
many equivalent ways, almost all of which are preferable to direct
quantization
of the predictor coefficients themselves. The PARCOR coefficients
can be quantized with a nonuniform quantizer or transformed
with an inverse sine or hyperbolic tangent function to flatten their
statistical
distribution and then quantized with a fixed uniform quantizer.
Each coefficient can be allocated a number of bits contingent on its
importance in accurately representing the speech spectrum [131].
Atal [2] reported a system which used a 20thorder predictor and
quantized the resulting set of PARCOR coefficients after
transformation
with an inverse sine function. The number of bits per coefficient
ranged from 5 for each of the lowest two PARCORs down to 1
each for
the six highestorder PARCORs, yielding a total of 40 bits per
frame.
If the parameters are updated 100 times/s, then a total of 4000 bps
for
the shortdelay predictor information is required. To save bit rate,
it
is possible to update the shortdelay predictor 50 times/s for a
total
bit rate contribution of 2000 bps [2]. The long delay predictor has a
delay parameter M, which requires 7 bits to cover the range of
pitch
periods to be expected with 8 kHz sampling rate of the input. The
longdelay predictor in the system mentioned above used delays of
M − 1,M,M + 1 and three gains each quantized to 4 or 5 bit
accuracy.
This gave a total of 20 bits/frame and added 2000 bps to the
overall bit rate.
Vector Quantization for Predictor Quantization Another
approach to quantizing the predictor parameters is called vector
quantization,
or VQ [45]. The basic principle of vector quantization is
depicted in Figure 7.7. VQ can be applied in any context where a
set
of parameters naturally groups together into a vector (in the sense
of
7.3 ClosedLoop Coders 115
Distortion
Measure
Code Book
2B Vectors
Table
Lookup
Code Book
2B Vectors
v i i
vˆ k vˆ k
vˆ i
Encoder Decoder
Fig. 7.7 Vector quantization.
linear algebra).14 For example the predictor coefficients can be
represented
by vectors of predictor coefficients, PARCOR coefficients,
cepstrum
values, log area ratios, or line spectrum frequencies. In VQ, a
vector, v, to be quantized is compared exhaustively to a
“codebook”
populated with representative vectors of that type. The index i of
the
closest vector ˆvi to v, according to a prescribed distortion
measure,
is returned from the exhaustive search. This index i then represents
the quantized vector in the sense that if the codebook is known, the
corresponding quantized vector ˆvi can be looked up. If the
codebook
has 2B entries, the index i can be represented by a B bit number.
The design of a vector quantizer requires a training set consisting
of a large number (L) of examples of vectors that are drawn from
the
same distribution as vectors to be quantized later, along with a
distortion
measure for comparing two vectors. Using an iterative process,
2B codebook vectors are formed from the training set of vectors.
The
“training phase” is computationally intense, but need only be done
once, and it is facilitated by the LBG algorithm [70]. One
advantage of
VQ is that the distortion measure can be designed to sort the
vectors
into perceptually relevant prototype vectors. The resulting
codebook is
then used to quantize general test vectors. Typically a VQ training
set
consists of at least 10 times the ultimate number of codebook
vectors,
i.e., L ≥ 10 · 2B.
14 Independent quantization of individual speech samples or
individual predictor coefficients
is called scalar quantization.
116 Digital Speech Coding
The quantization process is also time consuming because a given
vector must be systematically compared with all the codebook
vectors
to determine which one it is closest to. This has led to numerous
innovations
in codebook structure that speed up the look up process [45].
Difference Signal Quantization It is necessary to code the
difference
signal in ADPCM. Since the difference signal has the same
sampling
rate as the input signal, it is desirable to use as few bits as possible
for the quantizer. If straightforward Bbit uniform quantization is
used,
the additional bit rate would be B · fs. This was the approach used
in
earlier studies by Noll [84] as mentioned in Section 7.3.3. In order
to
reduce the bit rate for coding the difference samples, some sort of
block
coding must be applied. One approach is to design a multibit
quantizer
that only operates on the largest samples of the difference signal.
Atal
and Schroeder [2, 7] proposed to precede the quantizer by a “center
clipper,” which is a system that sets to zero value all samples
whose
amplitudes are within a threshold band and passes those samples
above
threshold unchanged. By adjusting the threshold, the number of
zero
samples can be controlled. For high thresholds, long runs of zero
samples
result, and the entropy of the quantized difference signal becomes
less than one. The blocks of zero samples can be encoded
efficiently
using variablelengthtoblock codes yielding average bits/sample
on
the order of 0.7 [2]. For a sampling rate of 8 kHz, this coding
method
needs a total of 5600 bps for the difference signal. When this bit
rate
of the differential coder is added to the bit rate for quantizing the
predictor information (≈4000 bps) the total comes to
approximately
10,000 bps.
7.3.3.2 Quality vs. Bit Rate for ADPCM Coders
ADPCM systems normally operate with an 8 kHz sampling rate,
which
is a compromise aimed at providing adequate intelligibility and
moderate
bit rate for voice communication. With this sampling rate, the
speech signal must be bandlimited to somewhat less than 4 kHz by
a
lowpass filter. Adaptive prediction and quantization can easily
lower
the bit rate to 32 kbps with no degradation in perceived quality
(with
7.3 ClosedLoop Coders 117
64 kbps log PCM toll quality as a reference). Of course, raising the
bit rate for ADPCM above 64 kbps cannot improve the quality
over
logPCM because of the inherent frequency limitation. However,
the
bit rate can be lowered below 32 kbps with only modest distortion
until about 10 kbps. Below this value, quality degrades
significantly.
In order to achieve neartollquality at low bit rates, all the
techniques
discussed above and more must be brought into play. This
increases
the system complexity, although not unreasonably so for current
DSP
hardware. Thus, ADPCM is an attractive coding method when toll
quality is required at modest cost, and where adequate transmission
and/or storage capacity is available to support bit rates on the order
of 10 kbps.
7.3.4 AnalysisbySynthesis Coding
While ADPCM coding can produce excellent results at moderate
bit
rates, its performance is fundamentally constrained by the fact that
the
difference signal has the same sampling rate as the input signal.
The
center clipping quantizer produces a difference signal that can be
coded
efficiently. However, this approach is clearly not optimal since the
center
clipper throws away information in order to obtain a sparse
sequence.
What is needed is a way of creating an excitation signal for the
vocal
tract filter that is both efficient to code and also produces decoded
speech of high quality. This can be done within the same closed
loop
framework as ADPCM, but with some significant modifications.
7.3.4.1 Basic AnalysisbySynthesis Coding System
Figure 7.8 shows the block diagram of another class of closedloop
digital speech coders. These systems are called “analysisby
synthesis
coders” because the excitation is built up using an iterative process
to
produce a “synthetic” vocal tract filter output ˆx[n] that matches
the
input speech signal according to a perceptually weighted error
criterion.
As in the case of the most sophisticated ADPCM systems, the
operations of Figure 7.8 are carried out on blocks of speech
samples. In
particular, the difference, d[n], between the input, x[n], and the
output
of the vocal tract filter, ˆx[n], is filtered with a linear filter called
the
118 Digital Speech Coding
Perceptual
Weighting
W(z)
Excitation
Generator
Vocal Tract
Filter
H(z)
+
x[n] d[n] dˆ[n]
xˆ[n]

+ d′[n]
Fig. 7.8 Structure of analysisbysynthesis speech coders.
perceptual weighting filter, W(z). As the first step in coding a
block of
speech samples, both the vocal tract filter and the perceptual
weighting
filter are derived from a linear predictive analysis of the block.
Then,
the excitation signal is determined from the perceptually weighted
difference
signal d_[n] by an algorithm represented by the block labeled
“Excitation Generator.”
Note the similarity of Figure 7.8 to the core ADPCM diagram of
Figure 7.5. The perceptual weighting and excitation generator
inside
the dotted box play the role played by the quantizer in ADPCM,
where
an adaptive quantization algorithm operates on d[n] to produce a
quantized
difference signal ˆ d[n], which is the input to the vocal tract
system.
In ADPCM, the vocal tract model is in the same position in the
closedloop
system, but instead of the synthetic output ˆx[n], a signal ˜x[n]
predicted from ˆx[n] is subtracted from the input to form the
difference
signal. This is a key difference. In ADPCM, the synthetic output
differs
from the input x[n] by the quantization error. In analysisby
synthesis,
ˆx[n] = x[n] − d[n], i.e., the reconstruction error is −d[n], and a
perceptually
weighted version of that error is minimized in the meansquared
sense by the selection of the excitation ˆ d[n].
7.3.4.2 Perceptual Weighting of the Difference Signal
Since −d[n] is the error in the reconstructed signal, it is desirable
to
shape its spectrum to take advantage of perceptual masking effects.
In
ADPCM, this is accomplished by quantization noise feedback or
preemphasis/
deemphasis filtering. In analysisbysynthesis coding, the input
7.3 ClosedLoop Coders 119
to the vocal tract filter is determined so as to minimize the
perceptuallyweighted
error d_[n]. The weighting is implemented by linear filtering,
i.e., d_[n] = d[n] ∗ w[n] with the weighting filter usually defined in
terms
of the vocal tract filter as the linear system with system function
W(z) = A(z/α1)
A(z/α2)
= H(z/α2)
H(z/α1) . (7.14)
The poles of W(z) lie at the same angle but at α2 times the radii of
the poles of H(z), and the zeros of W(z) are at the same angles but
at
radii of α1 times the radii of the poles of H(z). If α1 > α2 the
frequency
response is like a “controlled” inverse filter for H(z), which is the
shape
desired. Figure 7.9 shows the frequency response of such a filter,
where
typical values of α1 = 0.9 and α2 = 0.4 are used in (7.14). Clearly,
this
filter tends to emphasize the high frequencies (where the vocal
tract
filter gain is low) and it deemphasizes the low frequencies in the
error
signal. Thus, it follows that the error will be distributed in
frequency so
that relatively more error occurs at low frequencies, where, in this
case,
such errors would be masked by the high amplitude low
frequencies.
By varying α1 and α2 in (7.14) the relative distribution of error can
be
adjusted.
0 1000 2000 3000 4000
−30
−20
−10
0
10
20
30
frequency in Hz,
log magnitude in dB
Fig. 7.9 Comparison of frequency responses of vocal tract filter
and perceptual weighting
filter in analysisbysynthesis coding.
120 Digital Speech Coding
7.3.4.3 Generating the Excitation Signal
Most analysisbysynthesis systems generate the excitation from a
finite
fixed collection of input components, which we designate here as
fγ[n]
for 0 ≤ n ≤ L − 1, where L is the excitation frame length and γ
ranges
over the finite set of components. The input is composed of a finite
sum
of scaled components selected from the given collection, i.e.,
ˆ d[n] =
_N
k=1
βkfγk [n]. (7.15)
The βks and the sequences fγk [n] are chosen to minimize
E =
L_−1
n=0
((x[n] − h[n] ∗ ˆ d[n]) ∗ w[n])2, (7.16)
where h[n] is the vocal tract impulse response and w[n] is the
impulse
response of the perceptual weighting filter with system function
(7.14).
Since the component sequences are assumed to be known at both
the
coder and decoder, the βs and γs are all that is needed to represent
the
input ˆ d[n]. It is very difficult to solve simultaneously for the
optimum βs
and γs that minimize (7.16). However, satisfactory results are
obtained
by solving for the component signals one at a time.
Starting with the assumption that the excitation signal is zero
during
the current excitation15 analysis frame (indexed 0 ≤ n ≤ L − 1), the
output in the current frame due to the excitation determined for
previous
frames is computed and denoted ˆx0[n] for 0 ≤ n ≤ L − 1. Normally
this would be a decaying signal that could be truncated after L
samples.
The error signal in the current frame at this initial stage of the
iterative
process would be d0[n] = x[n] − ˆx0[n], and the perceptually
weighted
error would be d_
0[n] = d0[n] ∗ w[n]. Now at the first iteration stage,
assume that we have determined which of the collection of input
components
fγ1 [n] will reduce the weighted meansquared error the most
and we have also determined the required gain β1. By
superposition,
ˆx1[n] = ˆx0[n] + β1fγ1 [n] ∗ h[n]. The weighted error at the first
iteration
is d_
1[n] = (d0[n] − β1fγ1 [n] ∗ h[n] ∗ w[n]), or if we define the
perceptually
weighted vocal tract impulse response as h_[n] = h[n] ∗ w[n]
15 Several analysis frames are usually included in one linear
predictive analysis frame.
7.3 ClosedLoop Coders 121
and invoke the distributive property of convolution, then the
weighted
error sequence is d_
1[n] = d_
0[n] − β1fγ1 [n] ∗ h_[n]. This process is continued
by finding the next component, subtracting it from the previously
computed residual error, and so forth. Generalizing to the kth
iteration,
the corresponding equation takes the form:
d
_
k[n] = d
_
k−1[n] − βky
_
k[n], (7.17)
where y_[n] = fγk [n] ∗ h_[n] is the output of the perceptually
weighted
vocal tract impulse response due to the input component fγk [n].
The
meansquared error at stage k of the iteration is defined as
Ek =
L_−1
n=0
(d
_
k[n])2 =
L_−1
n=0
(d
_
k−1[n] − βky
_
k[n])2, (7.18)
where it is assumed that d_
k−1[n] is the weighted difference signal that
remains after k − 1 steps of the process.
Assuming that γk is known, the value of βk that minimizes Ek in
(7.18) is
βk =
L_−1
n=0
d
_
k−1[n]y
_
k[n]
L_−1
n=0
(y
_[n])2
, (7.19)
and the corresponding minimum meansquared error is
E
(min)
k =
L_−1
n=0
(d
_
k−1[n])2 − β2
k
L_−1
n=0
(y
_
k[n])2. (7.20)
While we have assumed in the above discussion that fγk [n] and
βk are known, we have not discussed how they can be found.
Equation
(7.20) suggests that the meansquared error can be minimized by
maximizing
β2
k
L_−1
n=0
(y
_
k[n])2 =
_
L_−1
n=0
d
_
k−1[n]y
_
k[n]
2
L_−1
n=0
(y
_
k[n])2
, (7.21)
122 Digital Speech Coding
which is essentially the normalized crosscorrelation between the
new
input component and the weighted residual error d_
k−1[n]. An exhaustive
search through the collection of input components will determine
which component fγk [n] will maximize the quantity in (7.21) and
thus
reduce the meansquared error the most. Once this component is
found,
the corresponding βk can be found from (7.19), and the new error
sequence computed from (7.17).
After N iterations of this process, the complete set of component
input sequences and corresponding coefficients will have been
determined.
16 If desired, the set of input sequences so determined can be
assumed and a new set of βs can be found jointly by solving a set
of
N linear equations.
7.3.4.4 MultiPulse Excitation Linear Prediction (MPLP)
The procedure described in Section 7.3.4.3 is quite general since
the
only constraint was that the collection of component input
sequences be
finite. The first analysisbysynthesis coder was called multipulse
linear
predictive coding [4]. In this system, the component input
sequences are
simply isolated unit impulse sequences, i.e., fγ[n] = δ[n − γ], where
γ
is an integer such that 0 ≤ γ ≤ L − 1. The excitation sequence
derived
by the process described in Section 7.3.4.3 is therefore of the form:
ˆ d[n] =
_N
k=1
βkδ[n − γk], (7.22)
where N is usually on the order of 4–5 impulses in a 5 ms (40
samples
for fs = 8 kHz) excitation analysis frame.17 In this case, an
exhaustive
search at each stage can be achieved by computing just one
crosscorrelation
function and locating the impulse at the maximum of the
crosscorrelation. This is because the components y_
k[n] are all just
shifted versions of h_[n].
16 Alternatively, N need not be fixed in advance. The error
reduction process can be stopped
when E
(min)
k falls below a prescribed threshold.
17 It is advantageous to include two or more excitation analysis
frames in one linear predictive
analysis frame, which normally is of 20 ms (160 samples) duration.
7.3 ClosedLoop Coders 123
In a speech coder based on multipulse analysisbysynthesis, the
speech signal is represented by quantized versions of the prediction
coefficients, the pulse locations, and the pulse amplitudes. The
predictor
coefficients can be coded as discussed for ADPCM by encoding
any
of the alternative representations either by scalar or vector
quantization.
A number of different approaches have been devised for coding
the excitation parameters. These usually involve some sort of
differential
coding scheme. At bit rates on the order of 10–16 kbps multipulse
coding can approach toll quality; however, below about 10 kbps,
quality degrades rapidly due to the many parameters that must be
coded [13].
Regularpulse excitation (RPE) is a special case of multipulse
excitation
where it is easier to encode the impulse locations. Specifically
after determining the location γ1 of the first impulse, all other
impulses
are located at integer multiples of a fixed spacing on either side of
γ1.
Thus, only γ1, the integer multiples, and the gain constants must be
encoded. Using a preset spacing of 4 and an excitation analysis
window
of 40 samples yields high quality at about 10 kbps [67]. Because
the
impulse locations are fixed after the first iteration, RPE requires
significantly
less computation than fullblown multipulse excitation analysis.
One version of RPE is the basis for the 13 kbps digital coder in the
GSM mobile communications system.
7.3.4.5 CodeExcited Linear Prediction (CELP)
Since the major portion of the bit rate of an analysisbysynthesis
coder
lies in the excitation signal, it is not surprising that a great deal of
effort has gone into schemes for finding excitation signals that are
easier
to encode than multipulse excitation, but yet maintain high quality
of reproduction of the speech signal. The next major innovation
in the history of analysisbysynthesis systems was called code
excited
linear predictive coding or CELP [112].18 In this scheme, the
excitation
signal components are Gaussian random sequences stored in a
18 Earlier work by Stewart [123] used a residual codebook
(populated by the Lloyd algorithm
[71]) with a low complexity trellis search.
124 Digital Speech Coding
“codebook.”19 If the codebook contains 2M sequences, then a
given
sequence can be specified by an Mbit number. Typical codebook
sizes
are 256, 512 or 1024. Within the codebook, the sequence lengths
are
typically L = 40–60 samples (5–7.5 ms at 8 kHz sampling rate). If
N
sequences are used to form the excitation, then a total of M · N bits
will be required to code the sequences. Additional bits are still
required
to code the βks. With this method, the decoder must also have a
copy
of the analysis codebook so that the excitation can be regenerated
at
the decoder.
The main disadvantage of CELP is the high computational cost of
exhaustively searching the codebook at each stage of the error
minimization.
This is because each codebook sequence must be filtered
with the perceptually weighted impulse response before computing
the
cross correlation with the residual error sequence. Efficient
searching
schemes and structured codebooks have eased the computational
burden,
and modern DSP hardware can easily implement the computations
in real time.
The basic CELP framework has been applied in the development
of
numerous speech coders that operate with bit rates in the range
from
4800 to 16000 bps. The Federal Standard 1016 (FS1016) was
adopted
by the Department of Defense for use in secure voice transmission
at 4800 bps. Another system called VSELP (vector sum excited
linear
prediction), which uses multiple codebooks to achieve a bit rate of
8000 bps, was adopted in 1989 for North American Cellular
Systems.
The GSM half rate coder operating at 5600 bps is based on the IS
54
VSELP coder. Due to the block processing structure, CELP coders
can introduce more than 40 ms delay into communication systems.
The
ITU G.728 lowdelay CELP coder uses very short excitation
analysis
frames and backward linear prediction to achieve high quality at a
bit
rate of 16,000 bps with delays less than 2 ms. Still another
standardized
CELP system is the ITU G.729 conjugatestructure algebraic
CELP
(CSACELP) coder that is used in some mobile communication
systems.
19 While it may seem counterintuitive that the excitation could be
comprised of white noise
sequence, remember that, in fact, the object of linear prediction is
to produce just such
a signal.
7.4 OpenLoop Coders 125
7.3.4.6 LongDelay Predictors in AnalysisbySynthesis
Coders
Longdelay predictors have an interesting interpretation in
analysisbysynthesis
coding. If we maintain a memory of one or more frames of
previous values of the excitation signal, we can add a term β0 ˆ d[n
− γ0]
to (7.15). The long segment of the past excitation acts as a sort of
codebook where the Lsample codebook sequences overlap by L −
1
samples. The first step of the excitation computation would be to
compute
γ0 using (7.21) and then β0 using (7.19). For each value of γ0 to be
tested, this requires that the weighted vocal tract impulse response
be
convolved with Lsample segments of the past excitation starting
with
sample γ0. This can be done recursively to save computation. Then
the
component β0 ˆ d[n − γ0] can be subtracted from the initial error to
start
the iteration process in either the multipulse or the CELP
framework.
The additional bit rate for coding β0 and γ0 is often well
worthwhile,
and longdelay predictors are used in many of the standardized
coders
mentioned above. This incorporation of components of the past
history
of the excitation has been referred to as “selfexcitation” [104] or
the
use of an “adaptive codebook” [19].
7.4 OpenLoop Coders
ADPCM coding has not been useful below about 9600 bps,
multipulse
coding has similar limitations, and CELP coding has not been used
at
bit rates below about 4800 bits/s. While these closedloop systems
have
many attractive features, it has not been possible to generate
excitation
signals that can be coded at low bit rates and also produce high
quality
synthetic speech. For bit rates below 4800 bps, engineers have
turned
to the vocoder principles that were established decades ago. We
call
these systems openloop systems because they do not determine
the
excitation by a feedback process.
7.4.1 The TwoState Excitation Model
Figure 7.10 shows a source/system model for speech that is closely
related to physical models for speech production. As we have
discussed
126 Digital Speech Coding
Fig. 7.10 Twostateexcitation model for speech synthesis.
in the previous chapters, the excitation model can be very simple.
Unvoiced sounds are produced by exciting the system with white
noise,
and voiced sounds are produced by a periodic impulse train
excitation,
where the spacing between impulses is the pitch period, P0. We
shall
refer to this as the twostate excitation model. The slowlytime
varying
linear system models the combined effects of vocal tract
transmission,
radiation at the lips, and, in the case of voiced speech, the lowpass
frequency shaping of the glottal pulse. The V/UV
(voiced/unvoiced
excitation) switch produces the alternating voiced and unvoiced
segments
of speech, and the gain parameter G controls the level of the
filter output. When values for V/UV decision, G, P0, and the
parameters
of the linear system are supplied at periodic intervals (frames)
then the model becomes a speech synthesizer, as discussed in
Chapter
8. When the parameters of the model are estimated directly
from a speech signal, the combination of estimator and synthesizer
becomes a vocoder or, as we prefer, an openloop
analysis/synthesis
speech coder.
7.4.1.1 Pitch, Gain, and V/UV Detection
The fundamental frequency of voiced speech can range from well
below
100 Hz for lowpitched male speakers to over 250 Hz for high
pitched
voices of women and children. The fundamental frequency varies
slowly,
with time, more or less at the same rate as the vocal tract motions.
It
is common to estimate the fundamental frequency F0 or
equivalently,
7.4 OpenLoop Coders 127
the pitch period P0 = 1/F0 at a frame rate of about 50–100 times/s.
To do this, short segments of speech are analyzed to detect
periodicity
(signaling voiced speech) or aperiodicity (signaling unvoiced
speech).
One of the simplest, yet most effective approaches to pitch
detection,
operates directly on the time waveform by locating corresponding
peaks
and valleys in the waveform, and measuring the times between the
peaks [43]. The STACF is also useful for this purpose as illustrated
in Figure 4.8, which shows autocorrelation functions for both
voiced
speech, evincing a peak at the pitch period, and unvoiced speech,
which
shows no evidence of periodicity. In pitch detection applications of
the
shorttime autocorrelation, it is common to preprocess the speech
by
a spectrum flattening operation such as center clipping [96] or
inverse
filtering [77]. This preprocessing tends to enhance the peak at the
pitch
period for voiced speech while suppressing the local correlation
due to
formant resonances.
Another approach to pitch detection is suggested by Figure 5.5,
which shows the cepstra of segments of voiced and unvoiced
speech. In
this case, a strong peak in the expected pitch period range signals
voiced
speech, and the location of the peak is the pitch period. Similarly,
lack
of a peak in the expected range signals unvoiced speech [83].
Figure 5.6
shows a sequence of shorttime cepstra which moves from
unvoiced to
voiced speech in going from the bottom to the top of the figure.
The
time variation of the pitch period is evident in the upper plots.
The gain parameter G is also found by analysis of short segments
of speech. It should be chosen so that the shorttime energy of the
synthetic
output matches the shorttime energy of the input speech signal.
For this purpose, the autocorrelation function value φˆn[0] at lag 0
or
the cepstrum value cˆn[0] at quefrency 0 can be used to determine
the
energy of the segment of the input signal.
For digital coding applications, the pitch period, V/UV, and gain
must be quantized. Typical values are 7 bits for pitch period (P0 =
0
signals UV) and 5 bits for G. For a frame rate of 50 frames/s, this
totals
600 bps, which is well below the bit rate used to encode the
excitation
signal in closedloop coders such as ADPCM or CELP. Since the
vocal
tract filter can be coded as in ADPCM or CELP, much lower total
bit
128 Digital Speech Coding
rates are common in openloop systems. This comes at a large cost
in
quality of the synthetic speech output, however.
7.4.1.2 Vocal Tract System Estimation
The vocal tract system in the synthesizer of Figure 7.10 can take
many
forms. The primary methods that have been used have been
homomorphic
filtering and linear predictive analysis as discussed in Chapters 5
and 6, respectively.
The Homomorphic Vocoder Homomorphic filtering can be used
to extract a sequence of impulse responses from the sequence of
cepstra
that result from shorttime cepstrum analysis. Thus, one cepstrum
computation can yield both an estimate of pitch and the vocal tract
impulse response. In the original homomorphic vocoder, the
impulse
response was digitally coded by quantizing each cepstrum value
individually
(scalar quantization) [86]. The impulse response, reconstructed
from the quantized cepstrum at the synthesizer, is simply
convolved
with the excitation created from the quantized pitch, voicing, and
gain
information, i.e., s[n] = Ge[n] ∗ h[n].20 In a more recent
application of
homomorphic analysis in an analysisbysynthesis framework [21],
the
cepstrum values were coded using vector quantization, and the
excitation
derived by analysisbysynthesis as described in Section 7.3.4.3. In
still another approach to digital coding, homomorphic filtering was
used
to remove excitation effects in the shorttime spectrum and then
three
formant frequencies were estimated from the smoothed spectra.
Figure
5.6 shows an example of the formants estimated for voiced speech.
These formant frequencies were used to control the resonance
frequencies
of a synthesizer comprised of a cascade of secondorder section
IIR
digital filters [111]. Such a speech coder is called a formant
vocoder.
LPC Vocoder Linear predictive analysis can also be used to
estimate
the vocal tract system for an openloop coder with twostate
20 Care must be taken at frame boundaries. For example, the
impulse response can be
changed at the time a new pitch impulse occurs, and the resulting
output can overlap
into the next frame.
7.4 OpenLoop Coders 129
excitation [6]. In this case, the prediction coefficients can be coded
in
one of the many ways that we have already discussed. Such
analysis/
synthesis coders are called LPC vocoders. A vocoder of this type
was standardized by the Department of Defense as the Federal
Standard
FS1015. This system is also called LPC10 (or LPC10e) because
a 10thorder covariance linear predictive analysis is used to
estimate
the vocal tract system. The LPC10 system has a bit rate of 2400
bps
using a frame size of 22.5 ms with 12 bits/frame allocated to pitch,
voicing and gain and the remaining bits allocated to the vocal tract
filter coded as PARCOR coefficients.
7.4.2 ResidualExcited Linear Predictive Coding
In Section 7.2, we presented Figure 7.4 as motivation for the use of
the source/system model in digital speech coding. This figure
shows
an example of inverse filtering of the speech signal using a
prediction
error filter, where the prediction error (or residual) is significantly
smaller and less lowpass in nature. The speech signal can be
reconstructed
from the residual by passing it through the vocal tract system
H(z) = 1/A(z). None of the methods discussed so far attempt
to directly code the prediction error signal in an openloop manner.
ADPCM and analysisbysynthesis systems derive the excitation to
the
synthesis filter by a feedback process. The twostate model
attempts
to construct the excitation signal by direct analysis and
measurement
of the input speech signal. Systems that attempt to code the
residual
signal directly are called residualexcited linear predictive (RELP)
coders.
Direct coding of the residual faces the same problem faced in
ADPCM or CELP: the sampling rate is the same as that of the
input,
and accurate coding could require several bits per sample. Figure
7.11
shows a block diagram of a RELP coder [128]. In this system,
which is
quite similar to voiceexcited vocoders (VEV) [113, 130], the
problem
of reducing the bit rate of the residual signal is attacked by
reducing
its bandwidth to about 800 Hz, lowering the sampling rate, and
coding
the samples with adaptive quantization. Adaptive delta modulation
was used in [128], but APCM could be used if the sampling rate is
130 Digital Speech Coding
Inverse
Filter
A(z)
Lowpass
&
Decimator
Adaptive
Quantizer
Linear
Predictive
Analysis
x[n] r[n]
Adaptive
Quantizer
Decoder
Interpolator
& Spectrum
Flattener
Synthesis
Filter
H(z)
x[n]
d[n]
{α }
ˆ[n]
{α }
d[n]
Encoder
Decoder
Fig. 7.11 Residualexcited linear predictive (RELP) coder and
decoder.
lowered to 1600 Hz. The 800 Hz band is wide enough to contain
several
harmonics of the highest pitched voices. The reduced bandwidth
residual
is restored to full bandwidth prior to its use as an excitation signal
by a nonlinear spectrum flattening operation, which restores higher
harmonics of voiced speech. White noise is also added according
to an
empirically derived recipe. In the implementation of [128], the
sampling
rate of the input was 6.8 kHz and the total bit rate was 9600 kbps
with 6800 bps devoted to the residual signal. The quality achieved
at
this rate was not significantly better than the LPC vocoder with a
twostate excitation model. The principle advantage of this system
is
that no hard V/UV decision must be made, and no pitch detection
is
required. While this system did not become widely used, its basic
principles
can be found in subsequent openloop coders that have produced
much better speech quality at bit rates around 2400 bps.
7.4.3 Mixed Excitation Systems
While twostate excitation allows the bit rate to be quite low, the
quality
of the synthetic speech output leaves much to be desired. The
output
of such systems is often described as “buzzy,” and in many cases,
errors
in estimating pitch period or voicing decision cause the speech to
sound
unnatural if not unintelligible. The weaknesses of the twostate
model
7.4 OpenLoop Coders 131
for excitation spurred interest in a mixed excitation model where a
hard decision between V and UV is not required. Such a model
was
first proposed by Makhoul et al. [75] and greatly refined by
McCree
and Barnwell [80].
Figure 7.12 depicts the essential features of the mixedexcitation
linear predictive coder (MELP) proposed by McCree and Barnwell
[80]. This configuration was developed as the result of careful
experimentation, which focused onebyone on the sources of
distortions
manifest in the twostate excitation coder such as buzziness and
tonal distortions. The main feature is that impulse train excitation
and noise excitation are added instead of switched. Prior to their
addition,
they each pass through a multiband spectral shaping filter. The
gains in each of five bands are coordinated between the two filters
so
that the spectrum of e[n] is flat. This mixed excitation helps to
model
shorttime spectral effects such as “devoicing” of certain bands
during
voiced speech.21 In some situations, a “jitter” parameter ΔP is
invoked
to better model voicing transitions. Other important features of the
MELP system are lumped into the block labeled “Enhancing
Filters.”
This represents adaptive spectrum enhancement filters used to
enhance
formant regions,22 and a spectrally flat “pulse dispersion filter”
whose
Fig. 7.12 Mixedexcitation linear predictive (MELP) decoder.
21 Devoicing is evident in frame 9 in Figure 5.6.
22 Such filters are also used routinely in CELP coders and are
often referred to as post
filters.
132 Digital Speech Coding
purpose is to reduce “peakiness” due to the minimumphase nature
of
the linear predictive vocal tract system.
These modifications to the basic twostate excitation LPC vocoder
produce marked improvements in the quality of reproduction of the
speech signal. Several new parameters of the excitation must be
estimated
at analysis time and coded for transmission, but these add only
slightly to either the analysis computation or the bit rate [80]. The
MELP coder is said to produce speech quality at 2400 bps that is
comparable
to CELP coding at 4800 bps. In fact, its superior performance
led to a new Department of Defense standard in 1996 and
subsequently
to MILSTD3005 and NATO STANAG 4591, which operates at
2400,
1200, and 600 bps.
7.5 FrequencyDomain Coders
Although we have discussed a wide variety of digital speech
coding
systems, we have really only focused on a few of the most
important
examples that have current relevance. There is much room for
variation
within the general frameworks that we have identified. There is
one more class of coders that should be discussed. These are called
frequencydomain coders because they are based on the principle
of decomposing the speech signal into individual frequency bands.
Figure 7.13 depicts the general nature of this large class of coding
systems. On the far left of the diagram is a set (bank) of analysis
bandpass
filters. Each of these is followed by a downsampler by a factor
appropriate for the bandwidth of its bandpass input. On the far
right
in the diagram is a set of “bandpass interpolators” where each
interpolator
is composed of an upsampler (i.e., raising the sampling rate by
the same factor as the corresponding downsampling box) followed
by
a bandpass filter similar to the analysis filter. If the filters are
carefully
designed, and the outputs of the downsamplers are connected to
the
corresponding inputs of the bandpass interpolators, then it is
possible
for the output ˆx[n] to be virtually identical to the input x[n] [24].
Such perfect reconstruction filter bank systems are the basis for a
wideranging
class of speech and audio coders called subband coders [25].
When the outputs of the filter bank are quantized by some sort of
quantizer, the output ˆx[n] is not equal to the input, but by careful
7.5 FrequencyDomain Coders 133
Fig. 7.13 Subband coder and decoder for speech audio.
design of the quantizer the output can be indistinguishable from the
input perceptually. The goal with such coders is, of course, for the
total
composite bit rate23 to be as low as possible while maintaining
high
quality.
In contrast to the coders that we have already discussed, which
incorporated the speech production model into the quantization
process,
the filter bank structure incorporates important features of the
speech perception model. As discussed in Chapter 3, the basilar
membrane
effectively performs a frequency analysis of the sound impinging
on the eardrum. The coupling between points on the basilar
membrane
results in the masking effects that we mentioned in Chapter 3. This
masking was incorporated in some of the coding systems discussed
so
far, but only in a rudimentary manner.
To see how subband coders work, suppose first that the individual
channels are quantized independently. The quantizers can be any
quantization operator that preserves the waveform of the signal.
Typically,
adaptive PCM is used, but ADPCM could be used in principle.
Because the downsampled channel signals are full band at the
lower sampling rate, the quantization error affects all frequencies
in
23 The total bit rate will be the sum of the products of the sampling
rates of the downsampled
channel signals times the number of bits allocated to each channel.
134 Digital Speech Coding
that band, but the important point is that no other bands are
affected
by the errors in a given band. Just as important, a particular band
can be quantized according to its perceptual importance.
Furthermore,
bands with low energy can be encoded with correspondingly low
absolute error.
The simplest approach is to preassign a fixed number of bits to
each channel. In an early nonuniformly spaced five channel
system,
Crochiere et al. [25] found that subband coders operating at 16
kbps
and 9.6 kbps were preferred by listeners to ADPCM coders
operating
at 22 and 19 kbps, respectively.
Much more sophisticated quantization schemes are possible in the
configuration shown in Figure 7.13 if, as suggested by the dotted
boxes,
the quantization of the channels is done jointly. For example, if the
channel bandwidths are all the same so that the output samples of
the
downsamplers can be treated as a vector, vector quantization can
be
used effectively [23].
Another approach is to allocate bits among the channels
dynamically
according to a perceptual criterion. This is the basis for modern
audio coding standards such as the various MPEG standards. Such
systems
are based on the principles depicted in Figure 7.13. What is not
shown in that figure is additional analysis processing that is done
to
determine how to allocate the bits among the channels. This
typically
involves the computation of a finegrained spectrum analysis using
the
DFT. From this spectrum it is possible to determine which
frequencies
will mask other frequencies, and on the basis of this a threshold of
audibility is determined. Using this audibility threshold, bits are
allocated
among the channels in an iterative process so that as much of the
quantization error is inaudible as possible for the total bit budget.
The
details of perceptual audio coding are presented in the recent
textbook
by Spanias [120].
7.6 Evaluation of Coders
In this chapter, we have discussed a wide range of options for
digital
coding of speech signals. These systems rely heavily on the
techniques of
linear prediction, filter banks, and cepstrum analysis and also on
models
7.6 Evaluation of Coders 135
for speech production and speech perception. All this knowledge
and
technique is brought to bear on the problem of reducing the bit rate
of
the speech representation while maintaining high quality
reproduction
of the speech signal.
Throughout this chapter we have mentioned bit rates, complexity
of computation, processing delay, and quality as important
practical
dimensions of the speech coding problem. Most coding schemes
have
many operations and parameters that can be chosen to tradeoff
among
the important factors. In general, increasing the bit rate will
improve
quality of reproduction; however, it is important to note that for
many
systems, increasing the bit rate indefinitely does not necessarily
continue
to improve quality. For example, increasing the bit rate of an
openloop twostateexcitation LPC vocoder above about 2400 bps
does
not improve quality very much, but lowering the bit rate causes
noticeable
degradation. On the other hand, improved pitch detection and
source modeling can improve quality in an LPC vocoder as
witnessed
by the success of the MELP system, but this generally would come
with an increase in computation and processing delay.
In the final analysis, the choice of a speech coder will depend on
constraints imposed by the application such as cost of
coder/decoder,
available transmission capacity, robustness to transmission errors,
and quality requirements. Fortunately, there are many possibilities
to
choose from.
8
TexttoSpeech Synthesis Methods
In this chapter, we discuss systems whose goal is to convert
ordinary
text messages into intelligible and natural sounding synthetic
speech
so as to transmit information from a machine to a human user. In
other words, we will be concerned with computer simulation of the
upper part of the speech chain in Figure 1.2. Such systems are
often
referred to as texttospeech synthesis (or TTS) systems, and their
general
structure is illustrated in Figure 8.1. The input to the TTS system
is text and the output is synthetic speech. The two fundamental
processes performed by all TTS systems are text analysis (to
determine
the abstract underlying linguistic description of the speech) and
speech synthesis (to produce the speech sounds corresponding to
the
text input).
Fig. 8.1 Block diagram of general TTS system.
136
8.1 Text Analysis 137
8.1 Text Analysis
The text analysis module of Figure 8.1 must determine three things
from the input text string, namely:
(1) pronunciation of the text string: the text analysis process
must decide on the set of phonemes that is to be spoken,
the degree of stress at various points in speaking, the intonation
of the speech, and the duration of each of the sounds
in the utterance
(2) syntactic structure of the sentence to be spoken: the
text analysis process must determine where to place pauses,
what rate of speaking is most appropriate for the material
being spoken, and how much emphasis should be given to
individual words and phrases within the final spoken output
speech
(3) semantic focus and ambiguity resolution: the text analysis
process must resolve homographs (words that are spelled
alike but can be pronounced differently, depending on context),
and also must use rules to determine word etymology
to decide on how best to pronounce names and foreign words
and phrases.
Figure 8.2 shows more detail on how text analysis is performed.
The input to the analysis is plain English text. The first stage of
processing
does some basic text processing operations, including detecting
the structure of the document containing the text (e.g., email
message
versus paragraph of text from an encyclopedia article), normalizing
the
text (so as to determine how to pronounce words like proper names
or
homographs with multiple pronunciations), and finally performing
a
linguistic analysis to determine grammatical information about
words
and phrases within the text. The basic text processing benefits from
an online dictionary of word pronunciations along with rules for
determining
word etymology. The output of the basic text processing step
is tagged text, where the tags denote the linguistic properties of the
words of the input text string.
138 TexttoSpeech Synthesis Methods
Fig. 8.2 Components of the text analysis process.
8.1.1 Document Structure Detection
The document structure detection module seeks to determine the
location
of all punctuation marks in the text, and to decide their significance
with regard to the sentence and paragraph structure of the input
text.
For example, an end of sentence marker is usually a period, ., a
question
mark, ?, or an exclamation point, !. However this is not always the
case as in the sentence, “This car is 72.5 in. long” where there are
two
periods, neither of which denote the end of the sentence.
8.1.2 Text Normalization
Text normalization methods handle a range of text problems that
occur
in real applications of TTS systems, including how to handle
abbreviations
and acronyms as in the following sentences:
Example 1: “I live on Bourbon St. in St. Louis”
Example 2: “She worked for DEC in Maynard MA”
8.1 Text Analysis 139
where, in Example 1, the text “St.” is pronounced as street or saint,
depending on the context, and in Example 2 the acronym DEC can
be pronounced as either the word “deck” (the spoken acronym) or
the name of the company, i.e., Digital Equipment Corporation, but
is
virtually never pronounced as the letter sequence “D E C.”
Other examples of text normalization include number strings like
“1920” which can be pronounced as the year “nineteen twenty” or
the
number “one thousand, nine hundred, and twenty,” and dates,
times
currency, account numbers, etc. Thus the string “$10.50” should be
pronounced as “ten dollars and fifty cents” rather than as a
sequence
of characters.
One other important text normalization problem concerns the
pronunciation
of proper names, especially those from languages other than
English.
8.1.3 Linguistic Analysis
The third step in the basic text processing block of Figure 8.2 is a
linguistic analysis of the input text, with the goal of determining,
for
each word in the printed string, the following linguistic properties:
• the part of speech (POS) of the word
• the sense in which each word is used in the current context
• the location of phrases (or phrase groups) with a sentence (or
paragraph), i.e., where a pause in speaking might be appropriate
• the presence of anaphora (e.g., the use of a pronoun to refer
back to another word unit)
• the word (or words) on which emphasis are to be placed, for
prominence in the sentence
• the style of speaking, e.g., irate, emotional, relaxed, etc.
A conventional parser could be used as the basis of the linguistic
analysis
of the printed text, but typically a simple, shallow analysis is
performed
since most linguistic parsers are very slow.
140 TexttoSpeech Synthesis Methods
8.1.4 Phonetic Analysis
Ultimately, the tagged text obtained from the basic text processing
block of a TTS system has to be converted to a sequence of tagged
phones which describe both the sounds to be produced as well as
the
manner of speaking, both locally (emphasis) and globally
(speaking
style). The phonetic analysis block of Figure 8.2 provides the
processing
that enables the TTS system to perform this conversion, with the
help of a pronunciation dictionary. The way in which these steps
are
performed is as follows.
8.1.5 Homograph Disambiguation
The homograph disambiguation operation must resolve the correct
pronunciation
of each word in the input string that has more than one
pronunciation.
The basis for this is the context in which the word occurs.
One simple example of homograph disambiguation is seen in the
phrases
“an absent boy” versus the sentence “do you choose to absent
yourself.”
In the first phrase the word “absent” is an adjective and the accent
is
on the first syllable; in the second phrase the word “absent” is a
verb
and the accent is on the second syllable.
8.1.6 LettertoSound (LTS) Conversion
The second step of phonetic analysis is the process of grapheme
tophoneme
conversion, namely conversion from the text to (marked)
speech sounds. Although there are a variety of ways of performing
this
analysis, perhaps the most straightforward method is to rely on a
standard
pronunciation dictionary, along with a set of lettertosound rules
for words outside the dictionary.
Figure 8.3 shows the processing for a simple dictionary search for
word pronunciation. Each individual word in the text string is
searched
independently. First a “whole word” search is initiated to see if the
printed word exists, in its entirety, in the word dictionary. If so, the
conversion to sounds is straightforward and the dictionary search
begins
on the next word. If not, as is the case most often, the dictionary
search
attempts to find affixes (both prefixes and suffixes) and strips them
8.1 Text Analysis 141
Fig. 8.3 Block diagram of dictionary search for proper word
pronunciation.
from the word attempting to find the “root form” of the word, and
then does another “whole word” search. If the root form is not
present
in the dictionary, a set of lettertosound rules is used to determine
the
best pronunciation (usually based on etymology of the word) of the
root form of the word, again followed by the reattachment of
stripped
out affixes (including the case of no stripped out affixes).
8.1.7 Prosodic Analysis
The last step in the text analysis system of Figure 8.2 is prosodic
analysis
which provides the speech synthesizer with the complete set of
synthesis controls, namely the sequence of speech sounds, their
durations,
and an associated pitch contour (variation of fundamental
frequency
with time). The determination of the sequence of speech sounds
is mainly performed by the phonetic analysis step as outlined
above.
The assignment of duration and pitch contours is done by a set of
pitch
142 TexttoSpeech Synthesis Methods
and duration rules, along with a set of rules for assigning stress and
determining where appropriate pauses should be inserted so that
the
local and global speaking rates appear to be natural.
8.2 Evolution of Speech Synthesis Systems
A summary of the progress in speech synthesis, over the period
1962–
1997, is given in Figure 8.4. This figure shows that there have been
3 generations of speech synthesis systems. During the first
generation
(between 1962 and 1977) formant synthesis of phonemes using
a terminal analog synthesizer was the dominant technology using
rules which related the phonetic decomposition of the sentence to
formant frequency contours. The synthesis suffered from poor
intelligibility
and poor naturalness. The second generation of speech synthesis
methods (from 1977 to 1992) was based primarily on an LPC
representation of subword units such as diphones (half phones).
By
carefully modeling and representing diphone units via LPC
parameters,
it was shown that good intelligibility synthetic speech could
be reliably obtained from text input by concatenating the
appropriate
diphone units. Although the intelligibility improved dramatically
Fig. 8.4 Time line of progress in speech synthesis and TTS
systems.
8.2 Evolution of Speech Synthesis Systems 143
over first generation formant synthesis, the naturalness of the
synthetic
speech remained low due to the inability of single diphone units
to represent all possible combinations of sound using that diphone
unit. A detailed survey of progress in texttospeech conversion up
to 1987 is given in the review paper by Klatt [65]. The synthesis
examples that accompanied that paper are available for listening at
http://www.cs.indiana.edu/rhythmsp/ASA/Contents.html.
The third generation of speech synthesis technology was the period
from 1992 to the present, in which the method of “unit selection
synthesis”
was introduced and perfected, primarily by Sagisaka at ATR
Labs in Kyoto [108]. The resulting synthetic speech from this third
generation technology had good intelligibility and naturalness that
approached that of humangenerated speech. We begin our
discussion
of speech synthesis approaches with a review of early systems, and
then
discuss unit selection methods of speech synthesis later in this
chapter.
8.2.1 Early Speech Synthesis Approaches
Once the abstract underlying linguistic description of the text input
has
been determined via the steps of Figure 8.2, the remaining (major)
task
of TTS systems is to synthesize a speech waveform whose
intelligibility
is very high (to make the speech useful as a means of
communication
between a machine and a human), and whose naturalness is as
close to
real speech as possible. Both tasks, namely attaining high
intelligibility
along with reasonable naturalness, are difficult to achieve and
depend
critically on three issues in the processing of the speech synthesizer
“backend” of Figure 8.1, namely:
(1) choice of synthesis units: including whole words, phones,
diphones, dyads, or syllables
(2) choice of synthesis parameters: including LPC features,
formants, waveform templates, articulatory parameters, sinusoidal
parameters, etc.
(3) method of computation: including rulebased systems or
systems which rely on the concatenation of stored speech
units.
144 TexttoSpeech Synthesis Methods
8.2.2 Word Concatenation Synthesis
Perhaps the simplest approach to creating a speech utterance
corresponding
to a given text string is to literally splice together prerecorded
words corresponding to the desired utterance. For greatest
simplicity,
the words can be stored as sampled waveforms and simply
concatenated
in the correct sequence. This approach generally produces
intelligible,
but unnatural sounding speech, since it does not take into account
the
“coarticulation” effects of producing phonemes in continuous
speech,
and it does not provide either for the adjustment of phoneme
durations
or the imposition of a desired pitch variation across the utterance.
Words spoken in continuous speech sentences are generally much
shorter in duration than when spoken in isolation (often up to 50%
shorter) as illustrated in Figure 8.5. This figure shows wideband
spectrograms
for the sentence “This shirt is red” spoken as a sequence of
isolated words (with short, distinct pauses between words) as
shown
at the top of Figure 8.5, and as a continuous utterance, as shown at
the bottom of Figure 8.5. It can be seen that, even for this trivial
Fig. 8.5 Wideband spectrograms of a sentence spoken as a
sequence of isolated words (top
panel) and as a continuous speech utterance (bottom panel).
8.2 Evolution of Speech Synthesis Systems 145
example, the duration of the continuous sentence is on the order of
half
that of the isolated word version, and further the formant tracks of
the
continuous sentence do not look like a set of uniformly compressed
formant
tracks from the individual words. Furthermore, the boundaries
between words in the upper plot are sharply defined, while they are
merged in the lower plot.
The word concatenation approach can be made more sophisticated
by storing the vocabulary words in a parametric form (formants,
LPC
parameters, etc.) such as employed in the speech coders discussed
in
Chapter 7 [102]. The rational for this is that the parametric
representations,
being more closely related to the model for speech production,
can be manipulated so as to blend the words together, shorten
them, and impose a desired pitch variation. This requires that the
control
parameters for all the words in the task vocabulary (as obtained
from a training set of words) be stored as representations of the
words.
A special set of word concatenation rules is then used to create the
control signals for the synthesizer. Although such a synthesis
system
would appear to be an attractive alternative for general purpose
synthesis
of speech, in reality this type of synthesis is not a practical
approach.1
There are many reasons for this, but a major problem is that there
are far too many words to store in a word catalog for word
concatenation
synthesis to be practical except in highly restricted situations. For
example, there are about 1.7 million distinct surnames in the
United
States and each of them would have to be spoken and stored for a
general
word concatenation synthesis method. A second, equally important
limitation is that wordlength segments of speech are simply the
wrong
subunits. As we will see, shorter units such as phonemes or
diphones
are more suitable for synthesis.
Efforts to overcome the limitations of word concatenation followed
two separate paths. One approach was based on controlling the
motions
of a physical model of the speech articulators based on the
sequence of
phonemes from the text analysis. This requires sophisticated
control
1 It should be obvious that whole word concatenation synthesis
(from stored waveforms) is
also impractical for general purpose synthesis of speech.
146 TexttoSpeech Synthesis Methods
rules that are mostly derived empirically. From the vocal tract
shapes
and sources produced by the articulatory model, control parameters
(e.g., formants and pitch) can be derived by applying the acoustic
theory
of speech production and then used to control a synthesizer such
those
used for speech coding. An alternative approach eschews the
articulatory
model and proceeds directly to the computation of the control
signals
for a source/system model (e.g., formant parameters, LPC
parameters,
pitch period, etc.). Again, the rules (algorithm) for computing the
control
parameters are mostly derived by empirical means.
8.2.3 Articulatory Methods of Synthesis
Articulatory models of human speech are based upon the
application of
acoustic theory to physical models such as depicted in highly
stylized
form in Figure 2.1 [22]. Such models were thought to be inherently
more natural than vocal tract analog models since:
• we could impose known and fairly well understood physical
constraints on articular movements to create realistic
motions of the tongue, jaw, teeth, velum, etc.
• we could use Xray data (MRI data today) to study the
motion of the articulators in the production of individual
speech sounds, thereby increasing our understanding of the
dynamics of speech production
• we could model smooth articulatory parameter motions
between sounds, either via direct methods (namely solving
the wave equation), or indirectly by converting articulatory
shapes to formants or LPC parameters
• we could highly constrain the motions of the articulatory
parameters so that only natural motions would
occur, thereby potentially making the speech more natural
sounding.
What we have learned about articulatory modeling of speech is
that it requires a highly accurate model of the vocal cords and of
the
vocal tract for the resulting synthetic speech quality to be
considered
acceptable. It further requires rules for handling the dynamics of
the
8.2 Evolution of Speech Synthesis Systems 147
articulator motion in the context of the sounds being produced. So
far we have been unable to learn all such rules, and thus,
articulatory
speech synthesis methods have not been found to be practical for
synthesizing speech of acceptable quality.
8.2.4 Terminal Analog Synthesis of Speech
The alternative to articulatory synthesis is called terminal analog
speech
synthesis. In this approach, each sound of the language (phoneme)
is
characterized by a source excitation function and an ideal vocal
tract
model. Speech is produced by varying (in time) the excitation and
the
vocal tract model control parameters at a rate commensurate with
the
sounds being produced. This synthesis process has been called
terminal
analog synthesis because it is based on a model (analog) of the
human
vocal tract production of speech that seeks to produce a signal at its
output terminals that is equivalent to the signal produced by a
human
talker.2
The basis for terminal analog synthesis of speech is the
source/system model of speech production that we have used many
times in this text; namely an excitation source, e[n], and a transfer
function of the human vocal tract in the form of a rational system
function, i.e.,
H(z) = S(z)
E(z)
= B(z)
A(z)
=
b0 +
_q
k=1
bkz
−k
1 −
_p
k=1
akz
−k
, (8.1)
where S(z) is the ztransform of the output speech signal, s[n],
E(z) is the ztransform of the vocal tract excitation signal, e[n],
and {bk} = {b0,b1,b2, . . . , bq} and {ak} = {a1,a2, . . . , ap} are the
(timevarying)
coefficients of the vocal tract filter. (See Figures 2.2 and 4.1.)
2 In the designation “terminal analog synthesis”, the terms analog
and terminal result from
the historical context of early speech synthesis studies. This can be
confusing since today,
“analog” implies “not digital” as well as an analogous thing.
“Terminal” originally implied
the “output terminals” of an electronic analog (not digital) circuit
or system that was an
analog of the human speech production system.
148 TexttoSpeech Synthesis Methods
For most practical speech synthesis systems, both the excitation
signal
properties (P0 and V/UV ) and the filter coefficients of (8.1)
change
periodically so as to synthesize different phonemes.
The vocal tract representation of (8.1) can be implemented as a
speech synthesis system using a direct form implementation.
However,
it has been shown that it is preferable to factor the numerator and
denominator polynomials into either a series of cascade (serial)
resonances,
or into a parallel combination of resonances. It has also been
shown that an allpole model (B(z) =constant) is most appropriate
for
voiced (nonnasal) speech. For unvoiced speech a simpler model
(with
one complex pole and one complex zero), implemented via a
parallel
branch is adequate. Finally a fixed spectral compensation network
can be used to model the combined effects of glottal pulse shape
and
radiation characteristics from the lips and mouth, based on two real
poles in the zplane. A complete serial terminal analog speech
synthesis
model, based on the above discussion, is shown in Figure 8.6
[95, 111].
The voiced speech (upper) branch includes an impulse generator
(controlled by a timevarying pitch period, P0), a timevarying
voiced
signal gain, AV , and an allpole discretetime system that consists
of a cascade of 3 timevarying resonances (the first three formants,
F1,F2,F3), and one fixed resonance F4.
Fig. 8.6 Speech synthesizer based on a cascade/serial (formant)
synthesis model.
8.3 Unit Selection Methods 149
The unvoiced speech (lower) branch includes a white noise
generator, a timevarying unvoiced signal gain, AN, and a
resonance/
antiresonance system consisting of a timevarying pole (FP ) and
a timevarying zero (FZ).
The voiced and unvoiced components are added and processed by
the fixed spectral compensation network to provide the final
synthetic
speech output.
The resulting quality of the synthetic speech produced using a
terminal
analog synthesizer of the type shown in Figure 8.6 is highly
variable
with explicit model shortcomings due to the following:
• voiced fricatives are not handled properly since their mixed
excitation is not part of the model of Figure 8.6
• nasal sounds are not handled properly since nasal zeros are
not included in the model
• stop consonants are not handled properly since there is no
precise timing and control of the complex excitation signal
• use of a fixed pitch pulse shape, independent of the pitch
period, is inadequate and produces buzzy sounding voiced
speech
• the spectral compensation model is inaccurate and does not
work well for unvoiced sounds.
Many of the short comings of the model of Figure 8.6 are
alleviated by
a more complex model proposed by Klatt [64]. However, even
with a
more sophisticated synthesis model, it remains a very challenging
task
to compute the synthesis parameters. Nevertheless Klatt’s Klattalk
system achieved adequate quality by 1983 to justify
commercialization
by the Digital Equipment Corporation as the DECtalk system.
Some DECtalk systems are still in operation today as legacy
systems,
although the unit selection methods to be discussed in Section 8.3
now
provide superior quality in current applications.
8.3 Unit Selection Methods
The key idea of a concatenative TTS system, using unit selection
methods, is to use synthesis segments that are sections of
prerecorded
150 TexttoSpeech Synthesis Methods
natural speech [31, 50, 108]. The word concatenation method
discussed
in Section 8.2.2 is perhaps the simplest embodiment of this idea;
however,
as we discussed, shorter segments are required to achieve better
quality synthesis. The basic idea is that the more segments
recorded,
annotated, and saved in the database, the better the potential
quality
of the resulting speech synthesis. Ultimately, if an infinite number
of
segments were recorded and saved, the resulting synthetic speech
would
sound natural for virtually all possible synthesis tasks.
Concatenative
speech synthesis systems, based on unit selection methods, are
what is
conventionally known as “data driven” approaches since their
performance
tends to get better the more data that is used for training the
system and selecting appropriate units.
In order to design and build a unit selection system based on using
recorded speech segments, several issues have to be resolved,
including
the following:
(1) What speech units should be used as the basic synthesis
building blocks?
(2) How the synthesis units are selected (extracted) from natural
speech utterances?
(3) How the units are labeled for retrieval from a large database
of units?
(4) What signal representation should be used to represent the
units for storage and reproduction purposes?
(5) What signal processing methods can be used to spectrally
smooth the units (at unit junctures) and for prosody modification
(pitch, duration, amplitude)?
We now attempt to answer each of these questions.
8.3.1 Choice of Concatenation Units
The units for unit selection can (in theory) be as large as words and
as small as phoneme units. Words are prohibitive since there are
essentially
an infinite number of words in English. Subword units include
syllables (about 10,000 in English), phonemes (about 45 in
English,
but they are highly context dependent), demisyllables (about 2500
8.3 Unit Selection Methods 151
in English), and diphones (about 1500–2500 in English). The ideal
synthesis unit is context independent and easily concatenates with
other (appropriate) subword units [15]. Based on this criterion, the
most reasonable choice for unit selection synthesis is the set of
diphone
units.3
Before any synthesis can be done, it is necessary to prepare an
inventory
of units (diphones). This requires significant effort if done
manually.
In any case, a large corpus of speech must be obtained from which
to extract the diphone units as waveform snippets. These units are
then
represented in some compressed form for efficient storage. High
quality
coding such as MPLP or CELP is used to limit compression
artifacts.
At the final synthesis stage, the diphones are decoded into
waveforms
for final merging, duration adjustment and pitch modification, all
of
which take place on the time waveform.
8.3.2 From Text to Diphones
The synthesis procedure from subword units is straightforward, but
far
from trivial [14]. Following the process of text analysis into
phonemes
and prosody, the phoneme sequence is first converted to the
appropriate
sequence of units from the inventory. For example, the phrase “I
want”
would be converted to diphone units as follows:
Text Input: I want.
Phonemes: /#/ AY/ /W/ /AA/ /N/ /T/ /#/
Diphones: /# AY/ /AYW/ /WAA/ /AAN/ /NT/ /T #/,
where the symbol # is used to represent silence (at the beginning
and
end of each sentence or phrase).
The second step in online synthesis is to select the most
appropriate
sequence of diphone units from the stored inventory. Since each
diphone unit occurs many times in the stored inventory, the
selection
3A diphone is simply the concatenation of two phonemes that are
allowed by the language
constraints to be sequentially contiguous in natural speech. For
example, a diphone based
upon the phonemes AY and W would be denoted AY−W where
the − denotes the joining
of the two phonemes.
152 TexttoSpeech Synthesis Methods
of the best sequence of diphone units involves solving a dynamic
programming
search for the sequence of units that minimizes a specified
cost function. The cost function generally is based on diphone
matches
at each of the boundaries between diphones, where the diphone
match
is defined in terms of spectral matching characteristics, pitch
matching
characteristics, and possibly phase matching characteristics.
8.3.3 Unit Selection Synthesis
The “Unit Selection” problem is basically one of having a given
set of
target features corresponding to the spoken text, and then
automatically
finding the sequence of units in the database that most closely
match these features. This problem is illustrated in Figure 8.7
which
shows a target feature set corresponding to the sequence of sounds
(phonemes for this trivial example) /HH/ /EH/ /L/ /OW/ from the
word /hello/.4 As shown, each of the phonemes in this word have
multiple
representations (units) in the inventory of sounds, having been
extracted from different phonemic environments. Hence there are
many
versions of /HH/ and many versions of /EH/ etc., as illustrated in
Figure 8.7. The task of the unit selection module is to choose one
of
each of the multiple representations of the sounds, with the goal
being
to minimize the total perceptual distance between segments of the
chosen
sequence, based on spectral, pitch and phase differences
throughout
the sounds and especially at the boundary between sounds. By
specifying
costs associated with each unit, both globally across the unit,
HH
EH L OW •••
HH
EH
L
OW
“Trained” Perceptual
Distance
HH
HH
HH EH
EHEH
EH
EH
L
L
L L
L
OW
OW
OW
Fig. 8.7 Illustration of basic process of unit selection. (After Dutoit
[31].)
4 If diphones were used, the sequence would be /HHEH/ /EH
L/ /LOW/.
8.3 Unit Selection Methods 153
Fig. 8.8 Online unit selection based on a Viterbi search through a
lattice of alternatives.
and locally at the unit boundaries with adjacent units, we can find
the sequence of units that best “join each other” in the sense of
minimizing
the accumulated distance across the sequence of units. This
optimal sequence (or equivalently the optimal path through the
combination
of all possible versions of each unit in the string) can be found
using a Viterbi search (dynamic programming) [38, 99]. This
dynamic
programming search process is illustrated in Figure 8.8 for a 3 unit
search. The Viterbi search effectively computes the cost of every
possible
path through the lattice and determines the path with the lowest
total cost, where the transitional costs (the arcs) reflect the cost
of concatenation of a pair of units based on acoustic distances, and
the nodes represent the target costs based on the linguistic identity
of
the unit.
Thus there are two costs associated with the Viterbi search, namely
a nodal cost based on the unit segmental distortion (USD) which is
defined as the difference between the desired spectral pattern of the
target (suitably defined) and that of the candidate unit, throughout
the unit, and the unit concatenative distortion (UCD) which is
defined
as the spectral (and/or pitch and/or phase) discontinuity across the
boundaries of the concatenated units. By way of example, consider
a target context of the word “cart” with target phonemes /K/ /AH/
/R/ /T/, and a source context of the phoneme /AH/ obtained from
the
154 TexttoSpeech Synthesis Methods
Fig. 8.9 Illustration of unit selection costs associated with a string
of target units and a
presumptive string of selected units.
source word “want” with source phonemes /W/ /AH/ /N/ /T/. The
USD distance would be the cost (specified analytically) between
the
sound /AH/ in the context /W/ /AH/ /N/ versus the desired sound
in the context /K/ /AH/ /R/.
Figure 8.9 illustrates concatenative synthesis for a given string of
target units (at the bottom of the figure) and a string of selected
units
from the unit inventory (shown at the top of the figure). We focus
our attention on Target Unit tj and Selected Units θj and θj+1.
Associated
with the match between units tj and θj is a USD Unit Cost
and associated with the sequence of selected units, θj and θj+1, is a
UCD concatenation cost. The total cost of an arbitrary string of N
selected units, Θ = {θ1,θ2, . . . , θN}, and the string of N target
units,
T = {t1, t2, . . . , tN} is defined as:
d(Θ,T) =
_N
j=1
du(θj, tj) +
N_−1
j=1
dt(θj,θj+1), (8.2)
where du(θj, tj) is the USD cost associated with matching target
unit tj
with selected unit θj and dt(θj,θj+1) is the UCD cost associated
with
concatenating the units θj and θj+1. It should be noted that when
selected units θj and θj+1 come from the same source sentence and
are adjacent units, the UCD cost goes to zero, as this is as natural a
concatenation as can exist in the database. Further, it is noted that
generally there is a small overlap region between concatenated
units.
The optimal path (corresponding to the optimal sequence of units)
can
be efficiently computed using a standard Viterbi search ([38, 99])
in
8.3 Unit Selection Methods 155
which the computational demand scales linearly with both the
number
of target and the number of concatenation units.
The concatenation cost between two units is essentially the spectral
(and/or phase and/or pitch) discontinuity across the boundary
between
the units, and is defined as:
dt(θj,θj+1) =
_p+2
k=1
wkCk(θj,θj+1), (8.3)
where p is the size of the spectral feature vector (typically p = 12
mfcc
coefficients (melfrequency cepstral coefficients as explained in
Section
5.6.3), often represented as a VQ codebook vector), and the extra
two features are log power and pitch. The weights, wk are chosen
during
the unit selection inventory creation phase and are optimized using
a trialanderror procedure. The concatenation cost essentially
measures
a spectral plus log energy plus pitch difference between the two
concatenated units at the boundary frames. Clearly the definition of
concatenation cost could be extended to more than a single
boundary
frame. Also, as stated earlier, the concatenation cost is defined to
be
zero (dt(θj,θj+1) = 0) whenever units θj and θj+1 are consecutive
(in
the database) since, by definition, there is no discontinuity in either
spectrum or pitch in this case. Although there are a variety of
choices
for measuring the spectral/log energy/pitch discontinuity at the
boundary,
a common cost function is the normalized meansquared error in
the feature parameters, namely:
Ck(θj,θj+1) =
[fθj
k (m) − fθj+1
k (1)]2
σ2
k
, (8.4)
where fθj
k (l) is the kth feature parameter of the lth frame of segment θj ,
m is the (normalized) duration of each segment, and σ2
k is the variance
of the kth feature vector component.
The USD or target costs are conceptually more difficult to
understand,
and, in practice, more difficult to instantiate. The USD cost
associated with units θj and tj is of the form:
du(θj, tj) =
_q
i=1
wt
iφi
%
Ti(fθj
i ),Ti(ftj
i )
&
, (8.5)
156 TexttoSpeech Synthesis Methods
where q is the number of features that specify the unit θj or tj ,
wt
i, i = 1,2, . . . , q is a trained set of target weights, and Ti(·) can be
either a continuous function (for a set of features, fi, such as
segmental
pitch, power or duration), or a set of integers (in the case of
categorical
features, fi, such as unit identity, phonetic class, position in
the syllable from which the unit was extracted, etc.). In the latter
case, φi can be looked up in a table of distances. Otherwise, the
local
distance function can be expressed as a simple quadratic distance
of
the form
φi
%
Ti(fθj
i ),Ti(ftj
i )
&
=
'
Ti(fθj
i ) − Ti(ftj
i )
(2
. (8.6)
The training of the weights, wt
i , is done offline. For each phoneme
in each phonetic class in the training speech database (which might
be
the entire recorded inventory), each exemplar of each unit is
treated
as a target and all others are treated as candidate units. Using this
training set, a leastsquares system of linear equations can be
derived
from which the weight vector can be solved. The details of the
weight
training methods are described by Schroeter [114].
The final step in the unit selection synthesis, having chosen the
optimal
sequence of units to match the target sentence, is to smooth/modify
the selected units at each of the boundary frames to better match
the spectra, pitch and phase at each unit boundary. Various
smoothing
methods based on the concepts of time domain harmonic scaling
(TDHS) [76] and pitch synchronous overlap add (PSOLA) [20, 31,
82]
have been proposed and optimized for such
smoothing/modification at
the boundaries between adjacent diphone units.
8.4 TTS Applications
Speech technology serves as a way of intelligently and efficiently
enabling humans to interact with machines with the goal of
bringing
down cost of service for an existing service capability, or
providing
new products and services that would be prohibitively expensive
without
the automation provided by a viable speech processing interactive
8.5 TTS Future Needs 157
system. Examples of existing services where TTS enables
significant
cost reduction are as follows:
• a dialog component for customer care applications
• a means for delivering text messages over an audio connection
(e.g., cellphone)
• a means of replacing expensive recorded Interactive Voice
Response system prompts (a service which is particularly
valuable when the prompts change often during the course
of the day, e.g., stock price quotations).
Similarly some examples of new products and services which are
enabled by a viable TTS technology are the following:
• locationbased services, e.g., alerting you to locations of
restaurants, gas stations, stores in the vicinity of your current
location
• providing information in cars (e.g., driving directions, traffic
reports)
• unified messaging (e.g., reading email and fax messages)
• voice portal providing voice access to webbased services
• ecommerce agents
• customized News, Stock Reports, Sports scores, etc.
• giving voice to small and embedded devices for reporting
information and alerts.
8.5 TTS Future Needs
Modern TTS systems have the capability of producing highly
intelligible,
and surprisingly natural speech utterances, so long as the utterance
is not too long, or too syntactically complicated, or too technical.
The
biggest problem with most TTS systems is that they have no idea
as
to how things should be said, but instead rely on the text analysis
for
emphasis, prosody, phrasing and all socalled suprasegmental
features
of the spoken utterance. The more TTS systems learn how to
produce
contextsensitive pronunciations of words (and phrases), the more
natural
sounding these systems will become. Hence, by way of example,
158 TexttoSpeech Synthesis Methods
the utterance “I gave the book to John” has at least three different
semantic interpretations, each with different emphasis on words in
the
utterance, i.e.,
I gave the book to John, i.e., not to Mary or Bob.
I gave the book to John, i.e., not the photos or the
apple.
I gave the book to John, i.e., I did it, not someone else.
The second future need of TTS systems is improvements in the
unit
selection process so as to better capture the target cost for
mismatch
between predicted unit specification (i.e., phoneme name, duration,
pitch, spectral properties) and actual features of a candidate
recorded
unit. Also needed in the unit selection process is a better spectral
distance
measure that incorporates human perception measures so as to
find the best sequence of units for a given utterance.
Finally, better signal processing would enable improved
compression
of the units database, thereby making the footprint of TTS systems
small enough to be usable in handheld and mobile devices.
9
Automatic Speech Recognition (ASR)
In this chapter, we examine the process of speech recognition by
machine, which is in essence the inverse of the texttospeech
problem.
The driving factor behind research in machine recognition of
speech
has been the potentially huge payoff of providing services where
humans
interact solely with machines, thereby eliminating the cost of live
agents
and significantly reducing the cost of providing services.
Interestingly,
as a side benefit, this process often provides users with a natural
and
convenient way of accessing information and services.
9.1 The Problem of Automatic Speech Recognition
The goal of an ASR system is to accurately and efficiently convert
a speech signal into a text message transcription of the spoken
words,
independent of the device used to record the speech (i.e., the
transducer
or microphone), the speaker’s accent, or the acoustic environment
in
which the speaker is located (e.g., quiet office, noisy room,
outdoors).
That is, the ultimate goal, which has not yet been achieved, is to
perform
as well as a human listener.
159
160 Automatic Speech Recognition (ASR)
Fig. 9.1 Conceptual model of speech production and speech
recognition processes.
A simple conceptual model of the speech generation and speech
recognition processes is given in Figure 9.1, which is a simplified
version
of the speech chain shown in Figure 1.2. It is assumed that the
speaker
intends to express some thought as part of a process of conversing
with another human or with a machine. To express that thought,
the
speaker must compose a linguistically meaningful sentence, W, in
the
form of a sequence of words (possibly with pauses and other
acoustic
events such as uh’s, um’s, er’s etc.). Once the words are chosen,
the
speaker sends appropriate control signals to the articulatory speech
organs which form a speech utterance whose sounds are those
required
to speak the desired sentence, resulting in the speech waveform
s[n].We
refer to the process of creating the speech waveform from the
speaker’s
intention as the Speaker Model since it reflects the speaker’s accent
and
choice of words to express a given thought or request. The
processing
steps of the Speech Recognizer are shown at the right side of
Figure 9.1
and consist of an acoustic processor which analyzes the speech
signal
and converts it into a set of acoustic (spectral, temporal) features,
X,
which efficiently characterize the speech sounds, followed by a
linguistic
decoding process which makes a best (maximum likelihood)
estimate of
the words of the spoken sentence, resulting in the recognized
sentence
ˆW
. This is in essence a digital simulation of the lower part of the
speech
chain diagram in Figure 1.2.
Figure 9.2 shows a more detailed block diagram of the overall
speech
recognition system. The input speech signal, s[n], is converted to
the
sequence of feature vectors, X = {x1,x2, . . . ,xT }, by the feature
analysis
block (also denoted spectral analysis). The feature vectors are
computed
on a framebyframe basis using the techniques discussed in the
earlier chapters. In particular, the mel frequency cepstrum
coefficients
9.2 Building a Speech Recognition System 161
Fig. 9.2 Block diagram of an overall speech recognition system.
are widely used to represent the shorttime spectral characteristics.
The
pattern classification block (also denoted as the decoding and
search
block) decodes the sequence of feature vectors into a symbolic
representation
that is the maximum likelihood string, ˆW that could have
produced the input sequence of feature vectors. The pattern
recognition
system uses a set of acoustic models (represented as hidden
Markov
models) and a word lexicon to provide the acoustic match score for
each
proposed string. Also, an Ngram language model is used to
compute
a language model score for each proposed word string. The final
block
in the process is a confidence scoring process (also denoted as an
utterance
verification block), which is used to provide a confidence score for
each individual word in the recognized string. Each of the
operations
in Figure 9.2 involves many details and, in some cases, extensive
digital
computation. The remainder of this chapter is an attempt to give
the
flavor of what is involved in each part of Figure 9.2.
9.2 Building a Speech Recognition System
The steps in building and evaluating a speech recognition system
are
the following:
(1) choose the feature set and the associated signal processing
for representing the properties of the speech signal over time
162 Automatic Speech Recognition (ASR)
(2) choose the recognition task, including the recognition word
vocabulary (the lexicon), the basic speech sounds to represent
the vocabulary (the speech units), the task syntax or
language model, and the task semantics (if any)
(3) train the set of speech acoustic and language models
(4) evaluate performance of the resulting speech recognition
system.
Each of these steps may involve many choices and can involve
significant
research and development effort. Some of the important issues are
summarized in this section.
9.2.1 Recognition Feature Set
There is no “standard” set of features for speech recognition.
Instead,
various combinations of acoustic, articulatory, and auditory
features
have been utilized in a range of speech recognition systems. The
most
popular acoustic features have been the (LPCderived) mel
frequency
cepstrum coefficients and their derivatives.
A block diagram of the signal processing used in most modern
large vocabulary speech recognition systems is shown in Figure
9.3.
The analog speech signal is sampled and quantized at rates
between
8000 up to 20,000 samples/s. A first order (highpass) preemphasis
network
(1 − αz−1) is used to compensate for the speech spectral falloff
at higher frequencies and approximates the inverse to the mouth
transmission
frequency response. The preemphasized signal is next blocked
into frames of N samples, with adjacent frames spaced M samples
apart. Typical values for N and M correspond to frames of duration
15–40 ms, with frame shifts of 10 ms being most common; hence
adjacent
frames overlap by 5–30 ms depending on the chosen values of N
and M. A Hamming window is applied to each frame prior to
spectral
analysis using either standard spectral analysis or LPC methods.
Following (optionally used) simple noise removal methods, the
spectral
coefficients are normalized and converted to melfrequency
cepstral
coefficients via standard analysis methods of the type discussed
in Chapter 5 [27]. Some type of cepstral bias removal is often used
prior to calculation of the first and second order cepstral
derivatives.
9.2 Building a Speech Recognition System 163
Fig. 9.3 Block diagram of feature extraction process for feature
vector consisting of mfcc
coefficients and their first and second derivatives.
Typically the resulting feature vector is the set of cepstral
coefficients,
and their first and secondorder derivatives. It is typical to use
about
13 mfcc coefficients, 13 firstorder cepstral derivative coefficients,
and
13 secondorder derivative cepstral coefficients, making a D = 39
size
feature vector.
9.2.2 The Recognition Task
Recognition tasks vary from simple word and phrase recognition
systems
to large vocabulary conversational interfaces to machines. For
example, using a digits vocabulary, the task could be recognition
of a
string of digits that forms a telephone number, or identification
code,
or a highly constrained sequence of digits that form a password.
9.2.3 Recognition Training
There are two aspects to training models for speech recognition,
namely
acoustic model training and language model training. Acoustic
model
164 Automatic Speech Recognition (ASR)
training requires recording each of the model units (whole words,
phonemes) in as many contexts as possible so that the statistical
learning method can create accurate distributions for each of the
model states. Acoustic training relies on accurately labeled
sequences of
speech utterances which are segmented according to the
transcription,
so that training of the acoustic models first involves segmenting
the
spoken strings into recognition model units (via either a Baum–
Welch
[10, 11] or Viterbi alignment method), and then using the
segmented
utterances to simultaneously build model distributions for each
state of
the vocabulary unit models. The resulting statistical models form
the
basis for the pattern recognition operations at the heart of the ASR
system. As discussed in Section 9.3, concepts such as Viterbi
search are
employed in the pattern recognition process as well as in training.
Language model training requires a sequence of text strings that
reflect the syntax of spoken utterances for the task at hand.
Generally
such text training sets are created automatically (based on a model
of
grammar for the recognition task) or by using existing text sources,
such as magazine and newspaper articles, or closed caption
transcripts
of television news broadcasts, etc. Other times, training sets for
language
models can be created from databases, e.g., valid strings of
telephone
numbers can be created from existing telephone directories.
9.2.4 Testing and Performance Evaluation
In order to improve the performance of any speech recognition
system,
there must be a reliable and statistically significant way of
evaluating
recognition system performance based on an independent test set
of
labeled utterances. Typically we measure word error rate and
sentence
(or task) error rate as a measure of recognizer performance. A brief
summary of performance evaluations across a range of ASR
applications
is given in Section 9.4.
9.3 The Decision Processes in ASR
The heart of any automatic speech recognition system is the pattern
classification and decision operations. In this section, we shall give
a
brief introduction to these important topics.
9.3 The Decision Processes in ASR 165
9.3.1 Mathematical Formulation of the ASR Problem
The problem of automatic speech recognition is represented as a
statistical
decision problem. Specifically it is formulated as a Bayes
maximum
a posteriori probability (MAP) decision process where we seek to
find
the word string ˆW (in the task language) that maximizes the a
posteriori
probability P(WX) of that string, given the measured feature
vector, X, i.e.,
ˆW
= argmax
W
P(WX). (9.1)
Using Bayes’ rule we can rewrite (9.1) in the form:
ˆW
= argmax
W
P(XW)P(W)
P(X) . (9.2)
Equation (9.2) shows that the calculation of the a posteriori
probability
is decomposed into two terms, one that defines the a priori
probability
of the word sequence, W, namely P(W), and the other that defines
the
likelihood that the word string, W, produced the feature vector, X,
namely P(XW). For all future calculations we disregard the
denominator
term, P(X), since it is independent of the word sequence W
which is being optimized. The term P(XW) is known as the
“acoustic
model” and is generally denoted as PA(XW) to emphasize the
acoustic
nature of this term. The term P(W) is known as the “language
model” and is generally denoted as PL(W) to emphasize the
linguistic
nature of this term. The probabilities associated with PA(XW) and
PL(W) are estimated or learned from a set of training data that have
been labeled by a knowledge source, usually a human expert,
where the
training set is as large as reasonably possible. The recognition
decoding
process of (9.2) is often written in the form of a 3step process, i.e.,
ˆW=
argmax
) *+W,
Step 3
PA(XW)
) *+ ,
Step 1
PL(W)
) *+ ,
Step 2
, (9.3)
where Step 1 is the computation of the probability associated with
the
acoustic model of the speech sounds in the sentence W, Step 2 is
the
computation of the probability associated with the linguistic model
of
the words in the utterance, and Step 3 is the computation
associated
166 Automatic Speech Recognition (ASR)
with the search through all valid sentences in the task language for
the
maximum likelihood sentence.
In order to be more explicit about the signal processing and
computations
associated with each of the three steps of (9.3) we need to
more explicit about the relationship between the feature vector, X,
and
the word sequence W. As discussed above, the feature vector, X, is
a
sequence of acoustic observations corresponding to each of T
frames of
the speech, of the form:
X = {x1,x2, . . . ,xT }, (9.4)
where the speech signal duration is T frames (i.e., T times the
frame
shift in ms) and each frame, xt, t = 1,2, . . . , T is an acoustic feature
vector of the form:
xt = (xt1,xt2, . . . , xtD) (9.5)
that characterizes the spectral/temporal properties of the speech
signal
at time t and D is the number of acoustic features in each frame.
Similarly we can express the optimally decoded word sequence, W,
as:
W = w1,w2,...,wM, (9.6)
where there are assumed to be exactly M words in the decoded
string.
9.3.2 The Hidden Markov Model
The most widely used method of building acoustic models (for
both
phonemes and words) is the use of a statistical characterization
known
as Hidden Markov Models (HMMs) [33, 69, 97, 98]. Figure 9.4
shows
a simple Q = 5state HMM for modeling a whole word. Each
HMM
state is characterized by a mixture density Gaussian distribution
that
characterizes the statistical behavior of the feature vectors within
the
states of the model [61, 62]. In addition to the statistical feature
densities
within states, the HMM is also characterized by an explicit set of
state transitions, aij , which specify the probability of making a
transition
from state i to state j at each frame, thereby defining the time
sequence of the feature vectors over the duration of the word.
Usually
the selftransitions, aii are large (close to 1.0), and the jump
transitions,
a12,a23,a34,a45, in the model, are small (close to 0).
9.3 The Decision Processes in ASR 167
Fig. 9.4 Wordbased, lefttoright, HMM with 5 states.
The complete HMM characterization of a Qstate word model (or a
subword unit like a phoneme model) is generally written as
λ(A,B,π)
with state transition matrix A = {aij ,1 ≤ i, j ≤ Q}, state observation
probability density, B = {bj(xt),1 ≤ j ≤ Q}, and initial state
distribution,
π = {πi,1 ≤ i ≤ Q} with π1 set to 1 for the “lefttoright” models
of the type shown in Figure 9.4.
In order to train the HMM (i.e., learn the optimal model
parameters)
for each word (or subword) unit, a labeled training set of
sentences
(transcribed into words and subword units) is used to guide
an efficient training procedure known as the Baum–Welch
algorithm
[10, 11].1 This algorithm aligns each of the various words (or
subword
units) with the spoken inputs and then estimates the appropriate
means, covariances and mixture gains for the distributions in
each model state. The Baum–Welch method is a hillclimbing
algorithm
and is iterated until a stable alignment of models and speech
is obtained. The details of the Baum–Welch procedure are beyond
the scope of this chapter but can be found in several references on
Speech Recognition methods [49, 99]. The heart of the training
procedure
for reestimating HMM model parameters using the Baum–
Welch procedure is shown in Figure 9.5. An initial HMM model is
used to begin the training process. The initial model can be
randomly
chosen or selected based on a priori knowledge of the model
1 The Baum–Welch algorithm is also widely referred to as the
forward–backward method.
168 Automatic Speech Recognition (ASR)
Fig. 9.5 The Baum–Welch training procedure based on a given
training set of utterances.
parameters. The iteration loop is a simple updating procedure for
computing
the forward and backward model probabilities based on an input
speech database (the training set of utterances) and then optimizing
the model parameters to give an updated HMM. This process is
iterated
until no further improvement in probabilities occurs with each new
iteration.
It is a simple matter to go from the HMM for a whole word,
as shown in Figure 9.4, to an HMM for a subword unit (such as a
phoneme) as shown in Figure 9.6. This simple 3state HMM is a
basic
subword unit model with an initial state representing the statistical
characteristics at the beginning of a sound, a middle state
representing
the heart of the sound, and an ending state representing the spectral
characteristics at the end of the sound. A word model is made
by concatenating the appropriate subword HMMs, as illustrated in
Fig. 9.6 Subwordbased HMM with 3 states.
9.3 The Decision Processes in ASR 169
Fig. 9.7 Wordbased HMM for the word /is/ created by
concatenating 3state subword
models for the subword units /ih/ and /z/.
Figure 9.7, which concatenates the 3state HMM for the sound /IH/
with the 3state HMM for the sound /Z/, giving the word model for
the word “is” (pronounced as /IH Z/). In general the composition
of
a word (from subword units) is specified in a word lexicon or
dictionary;
however once the word model has been built it can be used much
the same as whole word models for training and for evaluating
word
strings for maximizing the likelihood as part of the speech
recognition
process.
We are now ready to define the procedure for aligning a sequence
of M word models, w1,ww,...,wM with a sequence of feature
vectors,
X = {x1,x2, . . . ,xT }. The resulting alignment procedure is
illustrated
in Figure 9.8. We see the sequence of feature vectors along the
horizontal
axis and the concatenated sequence of word states along the
vertical
axis. An optimal alignment procedure determines the exact best
matching
sequence between word model states and feature vectors such that
the first feature vector, x1, aligns with the first state in the first
word
model, and the last feature vector, xT , aligns with the last state in
the
Mth word model. (For simplicity we show each word model as a 5
state
HMM in Figure 9.8, but clearly the alignment procedure works for
any
size model for any word, subject to the constraint that the total
number
of feature vectors, T, exceeds the total number of model states, so
that every state has at least a single feature vector associated with
that
state.) The procedure for obtaining the best alignment between
feature
vectors and model states is based on either using the Baum–Welch
statistical
alignment procedure (in which we evaluate the probability of
every alignment path and add them up to determine the probability
of
the word string), or a Viterbi alignment procedure [38, 132] for
which
170 Automatic Speech Recognition (ASR)
Fig. 9.8 Alignment of concatenated HMM word models with
acoustic feature vectors based
on either a Baum–Welch or Viterbi alignment procedure.
we determine the single best alignment path and use the probability
score along that path as the probability measure for the current
word
string. The utility of the alignment procedure of Figure 9.8 is based
on
the ease of evaluating the probability of any alignment path using
the
Baum–Welch or Viterbi procedures.
We now return to the mathematical formulation of the ASR
problem
and examine in more detail the three steps in the decoding
Equation (9.3).
9.3.3 Step 1 — Acoustic Modeling
The function of the acoustic modeling step (Step 1) is to assign
probabilities
to the acoustic realizations of a sequence of words, given the
observed acoustic vectors, i.e., we need to compute the probability
that
the acoustic vector sequence X = {x1,x2, . . . ,xT } came from the
word
sequence W = w1,w2,...,wM (assuming each word is represented as
an
HMM) and perform this computation for all possible word
sequences.
9.3 The Decision Processes in ASR 171
This calculation can be expressed as:
PA(XW) = PA({x1,x2, . . . ,xT }w1,w2,...,wM). (9.7)
If we make the assumption that each frame, xt, is aligned with
word i(t) and HMM model state j(t) via the function wi(t)
j(t) and if we
assume that each frame is independent of every other frame, we
can
express (9.7) as the product
PA(XW) =
_T
t=1
PA
!
xtwi(t)
j(t)
"
, (9.8)
where we associate each frame of X with a unique word and state,
wi(t)
j(t), in the word sequence. Further, we calculate the local
probability
PA
_
xtwi(t)
j(t)
_
given that we know the word from which frame t came.
The process of assigning individual speech frames to the
appropriate
word model in an utterance is based on an optimal alignment
process
between the concatenated sequence of word models and the
sequence
of feature vectors of the spoken input utterance being recognized.
This
alignment process is illustrated in Figure 9.9 which shows the set
of
T feature vectors (frames) along the horizontal axis, and the set of
M
words (and word model states) along the vertical axis. The optimal
segmentation of these feature vectors (frames) into the M words is
shown by the sequence of boxes, each of which corresponds to one
of
the words in the utterance and its set of optimally matching feature
vectors.
We assume that each word model is further decomposed into a set
of states which reflect the changing statistical properties of the
feature
vectors over time for the duration of the word. We assume that
each word is represented by an Nstate HMM model, and we
denote
the states as Sj,j = 1,2,...,N. Within each state of each word there
is a probability density that characterizes the statistical properties
of
the feature vectors in that state. We have seen in the previous
section
that the probability density of each state, and for each word, is
learned
during a training phase of the recognizer. Using a mixture of
Gaussian
densities to characterize the statistical distribution of the feature
vectors
in each state, j, of the word model, which we denote as bj(xt ),
172 Automatic Speech Recognition (ASR)
Fig. 9.9 Illustration of time alignment process between unknown
utterance feature vectors
and set of M concatenated word models.
we get a statebased probability density of the form:
bj(xt) =
_K
k=1
cjkN[xt,μjk ,Ujk ], (9.9)
where K is the number of mixture components in the density
function,
cjk is the weight of the kth mixture component in state j, with the
constraint cjk ≥ 0, and N is a Gaussian density function with mean
vector, μjk, for mixture k for state j, and covariance matrix, Ujk, for
mixture k in state j. The density constraints are:
_K
k=1
cjk = 1, 1 ≤ j ≤ N (9.10)
_ ∞
−∞
bj(xt )dxt = 1 , 1 ≤ j ≤ N. (9.11)
We now return to the issue of the calculation of the probability of
frame xt being associated with the j(t)th state of the i(t)th word in
9.3 The Decision Processes in ASR 173
the utterance, PA(xtwi(t)
j(t)), which is calculated as
PA
!
xtwi(t)
j(t)
"
= bi(t)
j(t)(xt), (9.12)
The computation of (9.12) is incomplete since we have ignored the
computation of the probability associated with the links between
word
states, and we have also not specified how to determine the within
word
state, j, in the alignment between a given word and a set of feature
vectors corresponding to that word. We come back to these issues
later
in this section.
The key point is that we assign probabilities to acoustic
realizations
of a sequence of words by using hidden Markov models of the
acoustic
feature vectors within words. Using an independent (and
orthographically
labeled) set of training data, we “train the system” and learn
the parameters of the best acoustic models for each word (or more
specifically for each sound that comprises each word). The
parameters,
according to the mixture model of (9.9) are, for each state of the
model,
the mixture weights, the mean vectors, and the covariance matrix.
Although we have been discussing acoustic models for whole
words,
it should be clear that for any reasonable size speech recognition
task,
it is impractical to create a separate acoustic model for every
possible
word in the vocabulary since each word would have to be spoken
in
every possible context in order to build a statistically reliable
model
of the density functions of (9.9). Even for modest size vocabularies
of about 1000 words, the amount of training data required for word
models is excessive.
The alternative to word models is to build acousticphonetic
models
for the 40 or so phonemes in the English language and construct
the
model for a word by concatenating (stringing together
sequentially)
the models for the constituent phones in the word (as represented
in a
word dictionary or lexicon). The use of such subword acoustic
phonetic
models poses no real difficulties in either training or when used to
build up word models and hence is the most widely used
representation
for building word models in a speech recognition system. Stateof
theart
systems use contextdependent phone models as the basic units of
recognition [99].
174 Automatic Speech Recognition (ASR)
9.3.4 Step 2 — The Language Model
The language model assigns probabilities to sequences of words,
based
on the likelihood of that sequence of words occurring in the
context
of the task being performed by the speech recognition system.
Hence
the probability of the text string W =“Call home” for a telephone
number identification task is zero since that string makes no sense
for
the specified task. There are many ways of building Language
Models
for specific tasks, including:
(1) statistical training from text databases transcribed from
taskspecific dialogs (a learning procedure)
(2) rulebased learning of the formal grammar associated with
the task
(3) enumerating, by hand, all valid text strings in the language
and assigning appropriate probability scores to each string.
The purpose of the language model, or grammar, is to enable the
computation of the a priori probability, PL(W), of a word string,
W,
consistent with the recognition task [59, 60, 106]. Perhaps the most
popular way of constructing the language model is through the use
of a statistical Ngram word grammar that is estimated from a large
training set of text utterances, either from the task at hand or from
a
generic database with applicability to a wide range of tasks. We
now
describe the way in which such a language model is built.
Assume we have a large text training set of wordlabeled
utterances.
(Such databases could include millions or even tens of millions of
text
sentences.) For every sentence in the training set, we have a text
file
that identifies the words in that sentence. If we make the
assumption
that the probability of a word in a sentence is conditioned on only
the previous N − 1 words, we have the basis for an Ngram
language
model. Thus we assume we can write the probability of the
sentence
W, according to an Ngram language model, as
PL(W) = PL(w1,w2,...,wM) (9.13)
=
M_
n=1
PL(wnwn−1,wn−2,...,wn−N+1), (9.14)
9.3 The Decision Processes in ASR 175
where the probability of a word occurring in the sentence only
depends
on the previous N − 1 words and we estimate this probability by
counting the relative frequencies of Ntuples of words in the
training
set. Thus, for example, to estimate word “trigram” probabilities
(i.e., the probability that a word wn was preceded by the pair of
words
(wn−1,wn−2)), we compute this quantity as
P(wnwn−1,wn−2) = C(wn−2,wn−1,wn)
C(wn−2,wn−1) , (9.15)
where C(wn−2,wn−1,wn) is the frequency count of the word triplet
(i.e., the trigram of words) consisting of (wn−2,wn−1,wn) as it
occurs
in the text training set, and C(wn−2,wn−1) is the frequency count
of
the word doublet (i.e., bigram of words) (wn−2,wn−1) as it occurs
in
the text training set.
9.3.5 Step 3 — The Search Problem
The third step in the Bayesian approach to automatic speech
recognition
is to search the space of all valid word sequences from the
language
model, to find the one with the maximum likelihood of having
been spoken.
The key problem is that the potential size of the search space can
be astronomically large (for large vocabularies and high average
word
branching factor language models), thereby taking inordinate
amounts
of computing power to solve by heuristic methods. Fortunately,
through
the use of methods from the field of Finite State Automata Theory,
Finite State Network (FSN) methods have evolved that reduce the
computational burden by orders of magnitude, thereby enabling
exact
maximum likelihood solutions in computationally feasible times,
even
for very large speech recognition problems [81].
The basic concept of a finite state network transducer is illustrated
in Figure 9.10 which shows a word pronunciation network for the
word
/data/. Each arc in the state diagram corresponds to a phoneme in
the
word pronunciation network, and the weight is an estimate of the
probability
that the arc is utilized in the pronunciation of the word in context.
We see that for the word /data/ there are four total pronunciations,
176 Automatic Speech Recognition (ASR)
Fig. 9.10 Word pronunciation transducer for four pronunciations of
the word /data/.
(After Mohri [81].)
namely (along with their (estimated) pronunciation probabilities):
(1) /D/ /EY/ /D/ /AX/ — probability of 0.32
(2) /D/ /EY/ /T/ /AX/ — probability of 0.08
(3) /D/ /AE/ /D/ /AX/ — probability of 0.48
(4) /D/ /AE/ /T/ /AX/ — probability of 0.12.
The combined FSN of the 4 pronunciations is a lot more efficient
than
using 4 separate enumerations of the word since all the arcs are
shared
among the 4 pronunciations and the total computation for the full
FSN
for the word /data/ is close to 1/4 the computation of the 4 variants
of the same word.
We can continue the process of creating efficient FSNs for each
word in the task vocabulary (the speech dictionary or lexicon), and
then combine word FSNs into sentence FSNs using the appropriate
language model. Further, we can carry the process down to the
level of
HMM phones and HMM states, making the process even more
efficient.
Ultimately we can compile a very large network of model states,
model
phones, model words, and even model phrases, into a much
smaller
network via the method of weighted finite state transducers
(WFST),
which combine the various representations of speech and language
and
optimize the resulting network to minimize the number of search
states
(and, equivalently, thereby minimize the amount of duplicate
computation).
A simple example of such a WFST network optimization is given
in Figure 9.11 [81].
Using the techniques of network combination (which include
network
composition, determinization, minimization, and weight pushing)
and network optimization, the WFST uses a unified mathematical
9.4 Representative Recognition Performance 177
Fig. 9.11 Use of WFSTs to compile a set of FSNs into a single
optimized network to
minimize redundancy in the network. (After Mohri [81].)
framework to efficiently compile a large network into a minimal
representation
that is readily searched using standard Viterbi decoding
methods [38]. Using these methods, an unoptimized network with
1022
states (the result of the cross product of model states, model
phones,
model words, and model phrases) was able to be compiled down to
a mathematically equivalent model with 108 states that was readily
searched for the optimum word string with no loss of performance
or
word accuracy.
9.4 Representative Recognition Performance
The block diagram of Figure 9.1 represents a wide range of
possibilities
for the implementation of automatic speech recognition. In Section
9.2.1, we suggest how the techniques of digital speech analysis
discussed in Chapters 4–6 can be applied to extract a sequence of
feature
vectors from the speech signal, and in Section 9.3 we describe the
most widely used statistical pattern recognition techniques that are
employed for mapping the sequence of feature vectors into a
sequence
of symbols or words.
Many variations on this general theme have been investigated over
the past 30 years or more, and as we mentioned before, testing and
evaluation is a major part of speech recognition research and
development.
Table 9.1 gives a summary of a range of the performance of
some speech recognition and natural language understanding
systems
178 Automatic Speech Recognition (ASR)
Table 9.1 Word error rates for a range of speech recognition
systems.
Corpus Type of Vocabulary Word
speech size error rate (%)
Connected digit strings
(TI Database)
Spontaneous 11 (0–9, oh) 0.3
Connected digit strings
(AT&T Mall
Recordings)
Spontaneous 11 (0–9, oh) 2.0
Connected digit strings
(AT&T HMIHY_c )
Conversational 11 (0–9, oh) 5.0
Resource management
(RM)
Read speech 1000 2.0
Airline travel information
system (ATIS)
Spontaneous 2500 2.5
North American business
(NAB & WSJ)
Read text 64,000 6.6
Broadcast News Narrated News 210,000 ≈15
Switchboard Telephone 45,000 ≈27
conversation
Callhome Telephone 28,000 ≈35
conversation
that have been developed so far. This table covers a range of
vocabulary
sizes, speaking styles, and application contexts [92].
It can be seen that for a vocabulary of 11 digits, the word error
rates are very low (0.3% for a very clean recording environment
for
the TI (Texas Instruments) connected digits database [68]), but
when
the digit strings are spoken in a noisy shopping mall environment
the
word error rate rises to 2.0% and when embedded within
conversational
speech (the AT&T HMIHY (How May I Help You) c system)
the word
error rate increases significantly to 5.0%, showing the lack of
robustness
of the recognition system to noise and other background
disturbances
[44]. Table 9.1 also shows the word error rates for a range of
DARPA
tasks ranging from
• read speech of commands and informational requests about
a naval ships database (the resource management system or
RM) with a 1000 word vocabulary and a word error rate of
2.0%
• spontaneous speech input for booking airlines travel (the airline
travel information system, or ATIS [133]) with a 2500
word vocabulary and a word error rate of 2.5%
9.5 Challenges in ASR Technology 179
• read text from a range of business magazines and newspapers
(the North American business task, or NAB) with a
vocabulary of 64,000 words and a word error rate of 6.6%
• narrated news broadcasts from a range of TV news providers
like CNBC, (the broadcast news task) with a 210,000 word
vocabulary and a word error rate of about 15%
• recorded live telephone conversations between two unrelated
individuals (the switchboard task [42]) with a vocabulary of
45,000 words and a word error rate of about 27%, and a
separate task for live telephone conversations between two
family members (the call home task) with a vocabulary of
28,000 words and a word error rate of about 35%.
9.5 Challenges in ASR Technology
So far, ASR systems fall far short of human speech perception in
all
but the simplest, most constrained tasks. Before ASR systems
become
ubiquitous in society, many improvements will be required in both
system performance and operational performance. In the system
area
we need large improvements in accuracy, efficiency, and
robustness in
order to utilize the technology for a wide range of tasks, on a wide
range of processors, and under a wide range of operating
conditions.
In the operational area we need better methods of detecting when a
person is speaking to a machine and isolating the spoken input
from
the background, we need to be able to handle users talking over the
voice prompts (socalled bargein conditions), we need more
reliable
and accurate utterance rejection methods so we can be sure that a
word
needs to be repeated when poorly recognized the first time, and
finally
we need better methods of confidence scoring of words, phrases,
and
even sentences so as to maintain an intelligent dialog with a
customer.
Conclusion
We have attempted to provide the reader with a broad overview of
the field of digital speech processing and to give some idea as to
the
remarkable progress that has been achieved over the past 4–5
decades.
Digital speech processing systems have permeated society in the
form of
cellular speech coders, synthesized speech response systems, and
speech
recognition and understanding systems that handle a wide range of
requests about airline flights, stock price quotations, specialized
help
desks etc.
There remain many difficult problems yet to be solved before
digital
speech processing will be considered a mature science and
technology.
Our basic understanding of the human articulatory system and how
the
various muscular controls come together in the production of
speech is
rudimentary at best, and our understanding of the processing of
speech
in the human brain is at an even lower level.
In spite of our shortfalls in understanding, we have been able to
create
remarkable speech processing systems whose performance
increases
at a steady pace. A firm understanding of the fundamentals of
acoustics,
linguistics, signal processing, and perception provide the tools for
180
Conclusion 181
building systems that work and can be used by the general public.
As
we increase our basic understanding of speech, the application
systems
will only improve and the pervasiveness of speech processing in
our
daily lives will increase dramatically, with the end result of
improving
the productivity in our work and home environments.
Acknowledgments
We wish to thank Professor Robert Gray, EditorinChief of Now
Publisher’s
Foundations and Trends in Signal Processing, for inviting us
to prepare this text. His patience, technical advice, and editorial
skill
were crucial at every stage of the writing. We also wish to thank
Abeer Alwan, Yariv Ephraim, Luciana Ferrer, Sadaoki Furui, and
Tom
Quatieri for their detailed and perceptive comments, which greatly
improved the final result. Of course, we are responsible for any
weaknesses
or inaccuracies that remain in this text.
182
References
[1] J. B. Allen and L. R. Rabiner, “A unified theory of shorttime
spectrum
analysis and synthesis,” Proceedings of IEEE, vol. 65, no. 11, pp.
1558–1564,
November 1977.
[2] B. S. Atal, “Predictive coding of speech at low bit rates,” IEEE
Transactions
on Communications, vol. COM30, no. 4, pp. 600–614, April
1982.
[3] B. S. Atal and S. L. Hanauer, “Speech analysis and synthesis
by linear prediction
of the speech wave,” Journal of the Acoustical Society of America,
vol. 50,
pp. 561–580, 1971.
[4] B. S. Atal and J. Remde, “A new model of LPC exitation for
producing
naturalsounding speech at low bit rates,” Proceedings of IEEE
ICASSP,
pp. 614–617, 1982.
[5] B. S. Atal and M. R. Schroeder, “Adaptive predictive coding of
speech signals,”
Bell System Technical Journal, vol. 49, pp. 1973–1986, October
1970.
[6] B. S. Atal and M. R. Schroeder, “Predictive coding of speech
signals and subjective
error criterion,” IEEE Transactions on Acoustics, Speech, and
Signal
Processing, vol. ASSP27, pp. 247–254, June 1979.
[7] B. S. Atal and M. R. Schroeder, “Improved quantizer for
adaptive predictive
coding of speech signals at low bit rates,” Proceedings of ICASSP,
pp. 535–538,
April 1980.
[8] T. B. Barnwell III, “Recursive windowing for generating
autocorrelation analysis
for LPC analysis,” IEEE Transactions on Acoustics, Speech and
Signal
Processing, vol. ASSP29, no. 5, pp. 1062–1066, October 1981.
[9] T. B. Barnwell III, K. Nayebi, and C. H. Richardson, Speech
Coding, A Computer
Laboratory Textbook. John Wiley and Sons, 1996.
183
184 References
[10] L. E. Baum, “An inequality and associated maximization
technique in statistical
estimation for probabilistic functions of Markov processes,”
Inequalities,
vol. 3, pp. 1–8, 1972.
[11] L. E. Baum, T. Petri, G. Soules, and N. Weiss, “A
maximization technique
occurring in the statistical analysis of probabilistic functions of
Markov
chains,” Annals in Mathematical Statistics, vol. 41, pp. 164–171,
1970.
[12] W. R. Bennett, “Spectra of quantized signals,” Bell System
Technical Journal,
vol. 27, pp. 446–472, July 1948.
[13] M. Berouti, H. Garten, P. Kabal, and P. Mermelstein,
“Efficient computation
and encoding of the multipulse excitation for LPC,” Proceedings
of ICASSP,
pp. 384–387, March 1984.
[14] M. Beutnagel, A. Conkie, and A. K. Syrdal, “Diphone
synthesis using
unit selection,” Third Speech Synthesis Workshop, Jenolan Caes,
Australia,
November 1998.
[15] M. Beutnatel and A. Conkie, “Interaction of units in a unit
selection
database,” Proceedings of Eurospeech ’99, Budapest, Hungary,
September
1999.
[16] B. P. Bogert, M. J. R. Healy, and J. W. Tukey, “The
quefrency alanysis of
times series for echos: Cepstrum, pseudoautocovariance, cross
cepstrum, and
saphe cracking,” in Proceedings of the Symposium on Time Series
Analysis,
(M. Rosenblatt, ed.), New York: John Wiley and Sons, Inc., 1963.
[17] E. Bresch, J. Nielsen, K. Nayak, and S. Narayanan,
“Synchornized and noiserobust
audio recordings during realtime MRI scans,” Journal of the
Acoustical
Society of America, vol. 120, no. 4, pp. 1791–1794, October 2006.
[18] C. S. Burrus and R. A. Gopinath, Introduction to Wavelets
and Wavelet Transforms.
PrenticeHall Inc., 1998.
[19] J. P. Campbell Jr., V. C. Welch, and T. E. Tremain, “An
expandable errorpretected
4800 bps CELP coder,” Proceedings of ICASSP, vol. 2, pp. 735–
738,
May 1989.
[20] F. Charpentier and M. G. Stella, “Diphone synthesis using an
overlapadd
technique for speech waveform concatenation,” Proceedings of
International
Conference on Acoustics, Speech and Signal Processing, pp.
2015–2018, 1986.
[21] J. H. Chung and R. W. Schafer, “Performance evaluation of
analysisbysynthesis
homomorphic vocoders,” Proceedings of IEEE ICASSP, vol. 2,
pp. 117–120, March 1992.
[22] C. H. Coker, “A model of articulatory dynamics and control,”
Proceedings of
IEEE, vol. 64, pp. 452–459, 1976.
[23] R. V. Cox, S. L. Gay, Y. Shoham, S. Quackenbush, N.
Seshadri, and N. Jayant,
“New directions in subband coding,” IEEE Journal of Selected
Areas in Communications,
vol. 6, no. 2, pp. 391–409, February 1988.
[24] R. E. Crochiere and L. R. Rabiner, Multirate Digital Signal
Processing.
PrenticeHall Inc., 1983.
[25] R. E. Crochiere, S. A. Webber, and J. L. Flanagan, “Digital
coding of speech
in subbands,” Bell System Technical Journal, vol. 55, no. 8, pp.
1069–1085,
October 1976.
References 185
[26] C. C. Cutler, “Differential quantization of communication
signals,” U.S.
Patent 2,605,361, July 29, 1952.
[27] S. B. Davis and P. Mermelstein, “Comparison of parametric
representations
for monosyllabic word recognition in continuously spoken
sentences,” IEEE
Transactions on Acoustics, Speech and Signal Processing, vol. 28,
pp. 357–366,
August 1980.
[28] F. deJager, “Delta modulation — a new method of PCM
transmission using
the 1unit code,” Philips Research Reports, pp. 442–466,
December 1952.
[29] P. B. Denes and E. N. Pinson, The speech chain. W. H.
Freeman Company,
2nd Edition, 1993.
[30] H. Dudley, “The vocoder,” Bell Labs Record, vol. 17, pp.
122–126, 1939.
[31] T. Dutoit, An Introduction to TexttoSpeech Synthesis.
Netherlands: Kluwer
Academic Publishers, 1997.
[32] G. Fant, Acoustic Theory of Speech Production. The Hague:
Mouton & Co.,
1960; Walter de Gruyter, 1970.
[33] J. D. Ferguson, “Hidden Markov Analysis: An Introduction,”
Hidden Markov
Models for Speech, Princeton: Institute for Defense Analyses,
1980.
[34] J. L. Flanagan, Speech Analysis, Synthesis and Perception.
SpringerVerlag,
1972.
[35] J. L. Flanagan, C. H. Coker, L. R. Rabiner, R. W. Schafer, and
N. Umeda,
“Synthetic voices for computers,” IEEE Spectrum, vol. 7, pp. 22–
45, October
1970.
[36] J. L. Flanagan, K. Ishizaka, and K. L. Shipley, “Synthesis of
speech from a
dynamic model of the vocal cords and vocal tract,” Bell System
Technical
Journal, vol. 54, no. 3, pp. 485–506, March 1975.
[37] H. Fletcher and W. J. Munson, “Loudness, its definition,
measurement and
calculation,” Journal of Acoustical Society of America, vol. 5, no.
2, pp. 82–
108, October 1933.
[38] G. D. Forney, “The Viterbi algorithm,” IEEE Proceedings,
vol. 61, pp. 268–
278, March 1973.
[39] S. Furui, “Cepstral analysis technique for automatic speaker
verification,”
IEEE Transactions on Acoustics Speech, and Signal Processing,
vol. ASSP
29, no. 2, pp. 254–272, April 1981.
[40] S. Furui, “Speaker independent isolated word recognition
using dynamic features
of speech spectrum,” IEEE Transactions on Acoustics, Speech,
Signal
Processing, vol. ASSP26, no. 1, pp. 52–59, February 1986.
[41] O. Ghitza, “Audiotry nerve representation as a basis for
speech processing,” in
Advances in Speech Signal Processing, (S. Furui and M. Sondhi,
eds.), pp. 453–
485, NY: Marcel Dekker, 1991.
[42] J. J. Godfrey, E. C. Holliman, and J. McDaniel,
“SWITCHBOARD: Telephone
Speech corpus for research and development,” Proceedings of
ICASSP 1992,
pp. 517–520, 1992.
[43] B. Gold and L. R. Rabiner, “Parallel processing techniques for
estimating
pitch period of speech in the time domain,” Journal of Acoustical
Society of
America, vol. 46, no. 2, pt. 2, pp. 442–448, August 1969.
186 References
[44] A. L. Gorin, B. A. Parker, R. M. Sachs, and J. G. Wilpon,
“How may I help
you?,” Proceedings of the Interactive Voice Technology for
Telecommunications
Applications (IVTTA), pp. 57–60, 1996.
[45] R. M. Gray, “Vector quantization,” IEEE Signal Processing
Magazine,
pp. 4–28, April 1984.
[46] R. M. Gray, “Toeplitz and circulant matrices: A review,”
Foundations and
Trends in Communications and Information Theory, vol. 2, no. 3,
pp. 155–239,
2006.
[47] J. A. Greefkes and K. Riemens, “Code modulation with
digitally controlled
companding for speech transmission,” Philips Technical Review,
pp. 335–353,
1970.
[48] H. Hermansky, “Auditory modeling in automatic recognition
of speech,” in
Proceedings of First European Conference on Signal Analysis and
Prediction,
pp. 17–21, Prague, Czech Republic, 1997.
[49] X. Huang, A. Acero, and H.W. Hon, Spoken Language
Processing. Prentice
Hall Inc., 2001.
[50] A. Hunt and A. Black, “Unit selection in a concatenative
speech synthesis
system using a large speech database,” Proceedings of ICASSP96,
Atlanta,
vol. 1, pp. 373–376, 1996.
[51] K. Ishizaka and J. L. Flanagan, “Synthesis of voiced sounds
from a twomass
model of the vocal cords,” Bell System Technical Journal, vol. 51,
no. 6,
pp. 1233–1268, 1972.
[52] F. Itakura, “Line spectrum representation of linear predictive
coefficients of
speech signals,” Journal of Acoustical Society of America, vol. 57,
pp. 535(a),
p. s35(A).
[53] F. Itakura and S. Saito, “Analysissynthesis telephony based
upon the maximum
likelihood method,” Proceedings of 6th International of Congress
on
Acoustics, pp. C17–C20, 1968.
[54] F. Itakura and S. Saito, “A statistical method for estimation of
speech spectral
density and formant frequencies,” Electronics and
Communications in Japan,
vol. 53A, no. 1, pp. 36–43, 1970.
[55] F. Itakura and T. Umezaki, “Distance measure for speech
recognition based on
the smoothed group delay spectrum,” in Proceedings of ICASSP87,
pp. 1257–
1260, Dallas TX, April 1987.
[56] N. S. Jayant, “Adaptive delta modulation with a onebit
memory,” Bell System
Technical Journal, pp. 321–342, March 1970.
[57] N. S. Jayant, “Adaptive quantization with one word memory,”
Bell System
Technical Journal, pp. 1119–1144, September 1973.
[58] N. S. Jayant and P. Noll, Digital Coding of Waveforms.
PrenticeHall, 1984.
[59] F. Jelinek, Statistical Methods for Speech Recognition.
Cambridge: MIT Press,
1997.
[60] F. Jelinek, R. L. Mercer, and S. Roucos, “Principles of lexical
language modeling
for speech recognition,” in Advances in Speech Signal Processing,
(S. Furui
and M. M. Sondhi, eds.), pp. 651–699, Marcel Dekker, 1991.
References 187
[61] B. H. Juang, “Maximum likelihood estimation for mixture
multivariate
stochastic observations of Markov chains,” AT&T Technology
Journal, vol. 64,
no. 6, pp. 1235–1249, 1985.
[62] B. H. Juang, S. E. Levinson, and M. M. Sondhi, “Maximum
likelihood estimation
for multivariate mixture observations of Markov chains,” IEEE
Transactions
in Information Theory, vol. 32, no. 2, pp. 307–309, 1986.
[63] B. H. Juang, L. R. Rabiner, and J. G. Wilpon, “On the use of
bandpass
liftering in speehc recognition,” IEEE Transactions on Acoustics,
Speech and
Signal Processing, vol. ASSP35, no. 7, pp. 947–954, July 1987.
[64] D. H. Klatt, “Software for a cascade/parallel formant
synthesizer,” Journal of
the Acoustical Society of America, vol. 67, pp. 971–995, 1980.
[65] D. H. Klatt, “Review of texttospeech conversion for
English,” Journal of the
Acoustical Society of America, vol. 82, pp. 737–793, September
1987.
[66] W. Koenig, H. K. Dunn, and L. Y. Lacey, “The sound
spectrograph,” Journal
of the Acoustical Society of America, vol. 18, pp. 19–49, 1946.
[67] P. Kroon, E. F. Deprettere, and R. J. Sluyter, “Regularpulse
excitation:
A nove approach to effective and efficient multipulse coding of
speech,”
IEEE Transactions on Acoustics, Speech, and Signal Processing,
vol. ASSP34,
pp. 1054–1063, October 1986.
[68] R. G. Leonard, “A database for speakerindependent digit
recognition,” Proceedings
of ICASSP 1984, pp. 42.11.1–42.11.4, 1984.
[69] S. E. Levinson, L. R. Rabiner, and M. M. Sondhi, “An
introduction to the
application of the theory of probabilistic functions of a Markov
process to
automatic speech recognition,” Bell System Technical Journal, vol.
62, no. 4,
pp. 1035–1074, 1983.
[70] Y. Linde, A. Buzo, and R. M. Gray, “An algorithm for vector
quantizer
design,” IEEE Transactions on Communications, vol. COM28,
pp. 84–95,
January 1980.
[71] S. P. Lloyd, “Least square quantization in PCM,” IEEE
Transactions on Information
Theory, vol. 28, pp. 129–137, March 1982.
[72] P. C. Loizou, Speech Enhancement, Theory and Practice.
CRC Press, 2007.
[73] R. F. Lyon, “A computational model of filtering, detection
and compression in
the cochlea,” in Proceedings of IEEE International Conference on
Acoustics,
Speech and Signal Processing, Paris, France, May 1982.
[74] J. Makhoul, “Linear prediction: A tutorial review,”
Proceedings of IEEE,
vol. 63, pp. 561–580, 1975.
[75] J. Makhoul, V. Viswanathan, R. Schwarz, and A. W. F.
Huggins, “A mixed
source model for speech compression and synthesis,” Journal of
the Acoustical
Society of America, vol. 64, pp. 1577–1581, December 1978.
[76] D. Malah, “Timedomain algorithms for harmonic bandwidth
reduction and
timescaling of pitch signals,” IEEE Transactions on Acoustics,
Speech and
Signal Processing, vol. 27, no. 2, pp. 121–133, 1979.
[77] J. D. Markel, “The SIFT algorithm for fundamental frequency
estimation,”
IEEE Transactions on Audio and Electroacoustics, vol. AU20, no.
5, pp. 367–
377, December 1972.
[78] J. D. Markel and A. H. Gray, Linear Prediction of Speech.
New York: Springer
Verlag, 1976.
188 References
[79] J. Max, “Quantizing for minimum distortion,” IRE
Transactions on Information
Theory, vol. IT6, pp. 7–12, March 1960.
[80] A. V. McCree and T. P. Barnwell III, “A mixed excitation
LPC vocoder
model for low bit rate speech coding,” IEEE Transactions on
Speech and
Audio Processing, vol. 3, no. 4, pp. 242–250, July 1995.
[81] M. Mohri, “Finitestate transducers in language and speech
processing,” Computational
Linguistics, vol. 23, no. 2, pp. 269–312, 1997.
[82] E. Moulines and F. Charpentier, “Pitch synchronous
waveform processing
techniques for texttospeech synthesis using diphones,” Speech
Communication,
vol. 9, no. 5–6, 1990.
[83] A. M. Noll, “Cepstrum pitch determination,” Journal of the
Acoustical Society
of America, vol. 41, no. 2, pp. 293–309, February 1967.
[84] P. Noll, “A comparative study of various schemes for speech
encoding,” Bell
System Technical Journal, vol. 54, no. 9, pp. 1597–1614,
November 1975.
[85] A. V. Oppenehim, “Superposition in a class of nonlinear
systems,” PhD dissertation,
MIT, 1964. Also: MIT Research Lab. of Electronics, Cambridge,
Massachusetts, Technical Report No. 432, 1965.
[86] A. V. Oppenheim, “A speech analysissynthesis system based
on homomorphic
filtering,” Journal of the Acoustical Society of America, vol. 45,
no. 2, pp. 293–
309, February 1969.
[87] A. V. Oppenheim, “Speech spectrograms using the fast
Fourier transform,”
IEEE Spectrum, vol. 7, pp. 57–62, August 1970.
[88] A. V. Oppenheim and R. W. Schafer, “Homomorphic analysis
of speech,”
IEEE Transactions on Audio and Electroacoustics, vol. AU16, pp.
221–228,
June 1968.
[89] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete
Time Signal Processing.
PrenticeHall Inc., 1999.
[90] A. V. Oppenheim, R. W. Schafer, and T. G. Stockham Jr.,
“Nonlinear filtering
of multiplied and convolved signals,” Proceedings of IEEE, vol.
56, no. 8,
pp. 1264–1291, August 1968.
[91] M. D. Paez and T. H. Glisson, “Minimum meansquared error
quantization in
speech,” IEEE Transactions on Communications, vol. Com20, pp.
225–230,
April 1972.
[92] D. S. Pallett et al., “The 1994 benchmark tests for the ARPA
spoken language
program,” Proceedings of 1995 ARPA Human Language
Technology Workshop,
pp. 5–36, 1995.
[93] M. R. Portnoff, “A quasionedimensional simulation for the
timevarying
vocal tract,” MS Thesis, MIT, Department of Electrical
Engineering, 1973.
[94] T. F. Quatieri, Discretetime speech signal processing.
Prentice Hall, 2002.
[95] L. R. Rabiner, “A model for synthesizing speech by rule,”
IEEE Transactions
on Audio and Electroacoustics, vol. AU17, no. 1, pp. 7–13, March
1969.
[96] L. R. Rabiner, “On the use of autocorrelation analysis for
pitch detection,”
IEEE Transactions on Acoustics, Speech, and Signal Processing,
vol. ASSP25,
no. 1, pp. 24–33, February 1977.
[97] L. R. Rabiner, “A tutorial on hidden Markov models and
selected applications
in speech recognition,” IEEE Proceedings, vol. 77, no. 2, pp. 257–
286, 1989.
References 189
[98] L. R. Rabiner and B. H. Juang, “An introduction to hidden
Markov models,”
IEEE Signal Processing Magazine, 1985.
[99] L. R. Rabiner and B. H. Juang, Fundamentals of Speech
Recognition. Prentice
Hall Inc., 1993.
[100] L. R. Rabiner and M. R. Sambur, “An algorithm for
determining the endpoints
of isolated utterances,” Bell System Technical Journal, vol. 54, no.
2, pp. 297–
315, February 1975.
[101] L. R. Rabiner and R. W. Schafer, Theory and Application of
Digital Speech
Processing. PrenticeHall Inc., 2009. (In preparation).
[102] L. R. Rabiner, R. W. Schafer, and J. L. Flanagan, “Computer
synthesis of
speech by concatenation of formant coded words,” Bell System
Technical Journal,
vol. 50, no. 5, pp. 1541–1558, May–June 1971.
[103] D. W. Robinson and R. S. Dadson, “A redetermination of
the equalloudness
contours for pure tones,” British Journal of Applied Physics, vol. 7,
pp. 166–
181, 1956.
[104] R. C. Rose and T. P. Barnwell III, “The self excited vocoder
— an alternate
approach to toll quality at 4800 bps,” Proceedings of ICASSP ’86,
vol. 11,
pp. 453–456, April 1986.
[105] A. E. Rosenberg, “Effect of glottal pulse shape on the quality
of natural vowels,”
Journal of the Acoustical Society of America, vol. 43, no. 4, pp.
822–828,
February 1971.
[106] R. Rosenfeld, “Two decades of statistical language
modeling: Where do we go
from here?,” IEEE Proceedings, vol. 88, no. 8, pp. 1270–1278,
2000.
[107] M. B. Sachs, C. C. Blackburn, and E. D. Young, “Rateplace
and temporalplace
representations of vowels in the auditory nerve and anteroventral
cochlear nucleus,” Journal of Phonetics, vol. 16, pp. 37–53, 1988.
[108] Y. Sagisaka, “Speech synthesis by rule using an optimal
selection of nonuniform
synthesis units,” Proceedings of the International Conference on
Acoustics, Speech and Signal Processing, pp. 679–682, 1988.
[109] R. W. Schafer, “Echo removal by discrete generalized linear
filtering,” PhD
dissertation, MIT, 1968. Also: MIT Research Laboratory of
Electronics, Cambridge,
Massachusetts, Technical Report No. 466, 1969.
[110] R. W. Schafer, “Homomorphic systems and cepstrum
analysis of speech,”
Springer Handbook of Speech Processing and Communication,
Springer, 2007.
[111] R. W. Schafer and L. R. Rabiner, “System for automatic
formant analysis of
voiced speech,” Journal of the Acoustical Society of America, vol.
47, no. 2,
pp. 458–465, February 1970.
[112] M. R. Schroeder and B. S. Atal, “Codeexcited linear
prediction (CELP):
Highquality speech at very low bit rates,” Proceedings of IEEE
ICASSP,
pp. 937–940, 1985.
[113] M. R. Schroeder and E. E. David, “A vocoder for
transmitting 10 kc/s speech
over a 3.5 kc/s channel,” Acustica, vol. 10, pp. 35–43, 1960.
[114] J. H. Schroeter, “Basic principles of speech synthesis,”
Springer Handbook of
Speech Processing, SpringerVerlag, 2006.
[115] S. Seneff, “A joint synchrony/meanrate model of auditory
speech processing,”
Journal of Phonetics, vol. 16, pp. 55–76, 1988.
190 References
[116] C. E. Shannon and W. Weaver, The Mathematical Theory of
Communication.
University of Illinois Press, Urbana, 1949.
[117] G. A. Sitton, C. S. Burrus, J. W. Fox, and S. Treitel,
“Factoring veryhighdegree
polynomials,” IEEE Signal Processing Magazine, vol. 20, no. 6,
pp. 27–
42, November 2003.
[118] B. Smith, “Instantaneous companding of quantized signals,”
Bell System Technical
Journal, vol. 36, no. 3, pp. 653–709, May 1957.
[119] F. K. Soong and B.H. Juang, “Optimal quantization of LSP
parameters,”
IEEE Transactions on Speech and Audio Processing, vol. 1, no. 1,
pp. 15–24,
January 1993.
[120] A. Spanias, T. Painter, and V. Atti, Audio Signal Processing
and Coding.
Wiley Interscience, 2007.
[121] K. N. Stevens, Acoustic Phonetics. MIT Press, 1998.
[122] S. S. Stevens and J. Volkman, “The relation of pitch to
frequency,” American
Journal of Psychology, vol. 53, p. 329, 1940.
[123] L. C. Stewart, R. M. Gray, and Y. Linde, “The design of
trellis waveform
coders,” IEEE transactions on Communications, vol. COM30, pp.
702–710,
April 1982.
[124] T. G. Stockham Jr., T. M. Cannon, and R. B. Ingebretsen,
“Blind deconvolution
through digital signal processing,” Proceedings of IEEE, vol. 63,
pp. 678–
692, April 1975.
[125] G. Strang and T. Nguyen, Wavelets and Filter Banks.
Wellesley, MA:
WellesleyCambridge Press, 1996.
[126] Y. Tohkura, “A weighted cepstral distance measure for
speech recognition,”
IEEE Transactions on Acoustics, Speech and Signal Processing,
vol. 35,
pp. 1414–1422, October 1987.
[127] J. M. Tribolet, “A new phase unwrapping algorithm,” IEEE
Transactions on
Acoustical, Speech, and Signal Processing, vol. ASSP25, no. 2,
pp. 170–177,
April 1977.
[128] C. K. Un and D. T. Magill, “The residualexcited linear
prediction vocoder
with transmission rate below 9.6 kbits/s,” IEEE Transactions on
Communications,
vol. COM23, no. 12, pp. 1466–1474, December 1975.
[129] P. P. Vaidyanathan, Multirate Systems and Filter Banks.
PrenticeHall Inc.,
1993.
[130] R. Viswanathan, W. Russell, and J. Makhoul, “Voiceexcited
LPC coders for
9.6 kbps speech transmission,” vol. 4, pp. 558–561, April 1979.
[131] V. Viswanathan and J. Makhoul, “Quantization properties of
transmission
parameters in linear predictive systems,” IEEE Transactions on
Acoustics,
Speech, and Signal Processing, vol. ASSP23, no. 3, pp. 309–321,
June 1975.
[132] A. J. Viterbi, “Error bounds for convolutional codes and an
asymptotically
optimal decoding Aalgorithm,” IEEE Transactions on Information
Theory,
vol. IT13, pp. 260–269, April 1967.
[133] W. Ward, “Evaluation of the CMU ATIS system,”
Proceedings of DARPA
Speech and Natural Language Workshop, pp. 101–105, February
1991.
[134] E. Zwicker and H. Fastl, Psychoacoustics. SpringerVerlag,
2nd Edition, 1990.
Supplemental References
The specific references that comprise the Bibliography of this text
are representative of the literature of the field of digital speech
processing.
In addition, we provide the following list of journals and
books as a guide for further study. Listing the books in
chronological
order of publication provides some perspective on the evolution of
the field.
Speech Processing Journals
• IEEE Transactions on Signal Processing. Main publication of
IEEE Signal
Processing Society.
• IEEE Transactions on Speech and Audio. Publication of IEEE
Signal
Processing Society that is focused on speech and audio processing.
• Journal of the Acoustical Society of America. General publication
of the
American Institute of Physics. Papers on speech and hearing as
well as
other areas of acoustics.
• Speech Communication. Published by Elsevier. A publication of
the
European Association for Signal Processing (EURASIP) and of the
International Speech Communication Association (ISCA).
191
192 Supplemental References
General Speech Processing References
• Speech Analysis, Synthesis and Perception, J. L. Flanagan,
Springer
Verlag, Second Edition, Berlin, 1972.
• Linear Prediction of Speech, J. D. Markel and A. H. Gray, Jr.,
Springer
Verlag, Berlin, 1976.
• Digital Processing of Speech Signals, L. R. Rabiner and R. W.
Schafer,
PrenticeHall Inc., 1978.
• Speech Analysis, R. W. Schafer and J. D. Markel (eds.), IEEE
Press
Selected Reprint Series, 1979.
• Speech Communication, Human and Machine, D.
O’Shaughnessy,
AddisonWesley, 1987.
• Advances in Speech Signal Processing, S. Furui and M. M.
Sondhi, Marcel
Dekker Inc., New York, 1991.
• DiscreteTime Processing of Speech Signals, J. Deller, Jr., J. H.
L.
Hansen, and J. G. Proakis, WileyIEEE Press, Classic Reissue,
1999.
• Acoustic Phonetics, K. N. Stevens, MIT Press, 1998.
• Speech and Audio Signal Processing, B. Gold and N. Morgan,
John Wiley
and Sons, 2000.
• Digital Speech Processing, Synthesis and Recognition, S. Furui,
Second
Edition, Marcel Dekker Inc., New York, 2001.
• DiscreteTime Speech Signal Processing, T. F. Quatieri, Prentice
Hall
Inc., 2002.
• Speech Processing, A Dynamic and OptimizationOriented
Approach,
L. Deng and D. O’Shaughnessy, Marcel Dekker, 2003.
• Springer Handbook of Speech Processing and Speech
Communication,
J. Benesty, M. M. Sondhi and Y Huang (eds.), Springer, 2008.
• Theory and Application of Digital Speech Processing, L. R.
Rabiner and
R. W. Schafer, Prentice Hall Inc., 2009.
Speech Coding References
• Digital Coding of Waveforms, N. S. Jayant and P. Noll, Prentice
Hall
Inc., 1984.
• Practical Approaches to Speech Coding, P. E. Papamichalis,
Prentice
Hall Inc., 1987.
• Vector Quantization and Signal Compression, A. Gersho and R.
M.
Gray, Kluwer Academic Publishers, 1992.
• Speech Coding and Synthesis, W. B. Kleijn and K. K. Paliwal,
Elsevier,
1995.
• Speech Coding, A Computer Laboratory Textbook, T. P.
Barnwell and
K. Nayebi, John Wiley and Sons, 1996.
Supplemental References 193
• A Practical Handbook of Speech Coders, R. Goldberg and L.
Riek, CRC
Press, 2000.
• Speech Coding Algorithms, W. C. Chu, John Wiley and Sons,
2003.
• Digital Speech: Coding for Low Bit Rate Communication
Systems, Second
Edition, A. M. Kondoz, John Wiley and Sons, 2004.
Speech Synthesis
• From Text to Speech, J. Allen, S. Hunnicutt and D. Klatt,
Cambridge
University Press, 1987.
• Acoustics of American English, J. P. Olive, A. Greenwood and J.
Coleman,
SpringerVerlag, 1993.
• Computing Prosody, Y. Sagisaka, N. Campbell and N. Higuchi,
Springer
Verlag, 1996.
• Progress in Speech Synthesis, J. VanSanten, R. W. Sproat, J. P.
Olive
and J. Hirschberg (eds.), SpringerVerlag, 1996.
• An Introduction to TexttoSpeech Synthesis, T. Dutoit, Kluwer
Academic
Publishers, 1997.
• Speech Processing and Synthesis Toolboxes, D. Childers, John
Wiley and
Sons, 1999.
• Text To Speech Synthesis: New Paradigms and Advances, S.
Narayanan
and A. Alwan (eds.), Prentice Hall Inc., 2004.
• TexttoSpeech Synthesis, P. Taylor, Cambridge University Press,
2008.
Speech Recognition and Natural Language Processing
• Fundamentals of Speech Recognition, L. R. Rabiner and B. H.
Juang,
Prentice Hall Inc., 1993.
• Connectionist Speech RecognitionA Hybrid Approach, H. A.
Bourlard
and N. Morgan, Kluwer Academic Publishers, 1994.
• Automatic Speech and Speaker Recognition, C. H. Lee, F. K.
Soong and
K. K. Paliwal (eds.), Kluwer Academic Publisher, 1996.
• Statistical Methods for Speech Recognition, F. Jelinek, MIT
Press, 1998.
• Foundations of Statistical Natural Language Processing, C. D.
Manning
and H. Schutze, MIT Press, 1999.
• Spoken Language Processing, X. Huang, A. Acero and H.W.
Hon, Prentice
Hall Inc., 2000.
• Speech and Language Processing, D. Jurafsky and J. H. Martin,
Prentice
Hall Inc., 2000.
• Mathematical Models for Speech Technology, S. E. Levinson,
John Wiley
and Sons, 2005.
194 Supplemental References
Speech Enhancement
• Digital Speech Transmission, Enhancement, Coding and Error
Concealment,
P. Vary and R. Martin, John Wiley and Sons, Ltd., 2006.
• Speech Enhancement, Theory and Practice, P. C. Loizou, CRC
Press,
2007.
Audio Processing
• Applications of Digital Signal Processing to Audio and
Acoustics,
H. Kahrs and K. Brandenburg (eds.), Kluwer Academic Publishers,
1998.
• Audio Signal Processing and Coding, A. Spanias, T. Painter and
V. Atti,
John Wiley and Sons, 2007.
variety of methods of representing speech in digital form, to applications in voice communication and automatic synthesis and recognition of speech. The breadth of this subject does not allow us to discuss any aspect of speech processing to great depth; hence our goal is to provide a useful introduction to the wide range of important concepts that comprise the field of digital speech processing. A more comprehensive treatment will appear in the forthcoming book, Theory and Application of Digital Speech Processing [101]. 1 Introduction The fundamental purpose of speech is communication, i.e., the transmission of messages. According to Shannon’s information theory [116], a message represented as a sequence of discrete symbols can be quantified by its information content in bits, and the rate of transmission of information is measured in bits/second (bps). In speech production, as well as in many humanengineered electronic communication systems, the information to be transmitted is encoded in the form of a continuously varying (analog) waveform that can be transmitted, recorded, manipulated, and ultimately decoded by a human listener. In the case of speech, the fundamental analog form of the message is an acoustic waveform, which we call the speech signal. Speech signals, as illustrated in Figure 1.1, can be converted to an electrical waveform by
a microphone, further manipulated by both analog and digital signal processing, and then converted back to acoustic form by a loudspeaker, a telephone handset or headphone, as desired. This form of speech processing is, of course, the basis for Bell’s telephone invention as well as today’s multitude of devices for recording, transmitting, and manipulating speech and audio signals. Although Bell made his invention without knowing the fundamentals of information theory, these ideas 3 4 Introduction 0 0.02 0.04 0.06 0.08 0.1 0.12 0.48 0.36 0.24 0.12 0 time in seconds SH UH D W IY CH EY S Fig. 1.1 A speech waveform with phonetic labels for the text message “Should we chase.” have assumed great importance in the design of sophisticated modern communications systems. Therefore, even though our main focus will be mostly on the speech waveform and its representation in the form of
parametric models, it is nevertheless useful to begin with a discussion of how information is encoded in the speech waveform. 1.1 The Speech Chain Figure 1.2 shows the complete process of producing and perceiving speech from the formulation of a message in the brain of a talker, to the creation of the speech signal, and finally to the understanding of the message by a listener. In their classic introduction to speech science, Denes and Pinson aptly referred to this process as the “speech chain” [29]. The process starts in the upper left as a message represented somehow in the brain of the speaker. The message information can be thought of as having a number of different representations during the process of speech production (the upper path in Figure 1.2). 1.1 The Speech Chain 5 Message Formulation Message Understanding Language Translation Language Code NeuroMuscular Controls Vocal Tract System Transmission Channel acoustic
waveform acoustic waveform Neural Transduction Basilar Membrane Motion discrete input continuous input Speech Production Speech Perception discrete output continuous output text phonemes, prosody articulatory motions semantics phonemes, words sentences feature extraction spectrum analysis 50 bps 200 bps 2000 bps 64700 Kbps information rate excitation, formants Fig. 1.2 The Speech Chain: from message, to speech signal, to understanding. For example the message could be represented initially as English text. In order to “speak” the message, the talker implicitly converts the text into a symbolic representation of the sequence of sounds corresponding to the spoken version of the text. This step, called the language code
the set of control signals that direct the neuromuscular system to move the speech articulators. The ARPAbet code does not require special fonts and is thus more convenient for computer applications.) The third step in the speech production process is the conversion to “neuromuscular controls.1 Thus. 1 The International Phonetic Association (IPA) provides a set of rules for phonetic transcription using an equivalent set of specialized symbols.” i. the text “should we chase” is represented phonetically (in ARPAbet symbols) as [SH UH D — W IY — CH EY S]. the speed and emphasis) in which the sounds are intended to be produced.e.2.e.. The end result of the neuromuscular controls step is a set of articulatory . namely the tongue. 6 Introduction jaw and velum. teeth. (See Chapter 2 for more discussion of phonetic transcription. the segments of the waveform of Figure 1. As an example.generator in Figure 1.1 are labeled with phonetic symbols using a computerkeyboardfriendly code called ARPAbet.. lips. in a manner that is consistent with the sounds of the desired spoken message and with the desired degree of emphasis. converts text symbols to phonetic symbols (along with stress and durational information) that describe the basic sounds of a spoken version of the message and the manner (i.
but if we include simple punctuation we get a count closer to 32 = 25 symbols). Furthermore.g. To determine the rate of information flow during speech production. At the second stage of the process. the rate of speaking for most people is about 10 symbols per second (somewhat on the high side. such as the one shown in Figure 1. Hence. the ARBAbet phonetic symbol set used to label the speech sounds in Figure 1. For example..1 contains approximately 64 = 26 symbols. pitch and stress) markers. that encodes the information in the desired message into the speech signal.1. but still acceptable for a rough information rate estimate). Finally the last step in the Speech Production process is the “vocal tract system” that physically creates the necessary sound sources and the appropriate vocal tract shapes over time so as to create an acoustic waveform. where the text representation is converted into phonemes and prosody (e. we estimate the base information rate of the text message as about 50 bps (5 bits per symbol times 10 symbols per second). assuming independent letters as a simple approximation. assume that there are about 32 symbols (letters) in the language (in English there are 26 letters. or about 6 bits/phoneme (again a rough approximation assuming independence . the information rate is estimated to increase by a factor of 4 to about 200 bps.motions (continuous control) that cause the vocal tract articulators to move in a prescribed manner in order to create the desired sounds.
loudness) could easily add 100 bps to the total information rate for a message encoded as a speech signal. the representation becomes continuous (in the form of control signals for articulatory motion). Additional information required to describe prosodic features of the signal (e. we could estimate the spectral bandwidth of these 1. duration. Thus..g. there are 8 phonemes in approximately 600 ms. If they could be measured.1. The articulators move relatively slowly compared to the time variation of the resulting acoustic waveform. pitch. The information representations for the first two stages in the speech chain are discrete so we can readily estimate the rate of information flow with some simple assumptions. the original text message is represented by a set of continuously varying signals whose digital representation requires a much higher data rate than .6 = 80 bps. Estimates of bandwidth and required accuracy suggest that the total data rate of the sampled articulatory control signals is about 2000 bps [34]. In Figure 1. This leads to an estimate of 8 × 6/0.of phonemes).1 The Speech Chain 7 control signals and appropriately sample and quantize these signals to obtain equivalent digital signals for which the data rate could be estimated. For the next stage in the speech production part of the speech chain.
the reproduced acoustic signal will be virtually indistinguishable from the original speech signal. We arrive at such numbers by examining the sampling rate and quantization required to represent the speech signal with a desired perceptual fidelity. the data rate of the digitized speech waveform at the end of the speech production part of the speech chain can be anywhere from 64. As we move from text to speech waveform through the speech chain. the result is an encoding of the message that can be effectively transmitted by acoustic wave propagation and robustly decoded by the hearing mechanism of a listener. it will sound different from the original speech signal uttered by the talker. resulting in a bit rate of 64. This representation is highly intelligible (i. For example.2 Finally. In this case. or a data rate of 705.000 bps.600 bps.000 to more than 700. Each sample can be quantized with 8 bits on a log scale.. “telephone quality” requires that a bandwidth of 0–4 kHz be preserved. humans can readily extract the message from it) but to most listeners. The above analysis of data rates shows that as we move from text to sampled speech waveform.the information rate that we estimated for transmission of the message as a speech signal. the speech waveform can be represented with “CD quality” using a sampling rate of 44. as we will see later. On the other hand.000 bps.100 samples/s with 16 bit samples. the data rate .e. implying a sampling rate of 8000 samples/s.
but much of it is due to the inefficiency 2 Note that we introduce the term data rate for digital representations to distinguish from the inherent information content of the message represented by the speech signal. of the type discussed above. a central theme of much of digital speech processing is to obtain a digital representation with lower data rate than that of the sampled waveform. The first step is the effective conversion of the acoustic waveform to a spectral representation.can increase by a factor of 10. speech mannerisms. which acts as a nonuniform spectrum analyzer by spatially separating the spectral components of the incoming speech signal and thereby analyzing them by what amounts to a nonuniform filter bank. as shown progressing to the left in the bottom half of Figure 1. as well as a speech perception/recognition model. The complete speech chain consists of a speech production/ generation model. motivated by an awareness of the low intrinsic information rate of speech. etc.000. Thus.. This is done within the inner ear by the basilar membrane. accent. The next step in the speech perception process is a neural transduction of the spectral features into a set of sound features (or distinctive features as they are . The speech perception model shows the series of steps from capturing speech at the ear to understanding the message encoded in the speech signal. 8 Introduction of simply sampling and finely quantizing analog signals.2. Part of this extra information represents characteristics of the talker such as emotional state.
this transmission channel consists of just the acoustic wave connection between 1.2 is rudimentary at best. and thus the entire model is useful for thinking about the processes that occur.referred to in the area of linguistics) that can be decoded and processed by the brain. words. Our fundamental understanding of the processes in most of the speech perception modules in Figure 1. The next step in the process is a conversion of the sound features into the set of phonemes. the last step in the speech perception model is the conversion of the phonemes. Finally. but it is generally agreed that some physical correlate of each of the steps in the speech perception model occur within the human brain. words and sentences of the message into an understanding of the meaning of the basic message in order to be able to respond to or take some appropriate action. There is one additional process shown in the diagram of the complete speech chain in Figure 1.2 that we have not discussed — namely the transmission channel between the speech generation and speech perception parts of the model. and sentences associated with the incoming message by a language translation process in the human brain. It is essential to include this transmission channel in our model for the speech chain since it includes real world noise and channel distortions that . In its simplest embodiment.2 Applications of Digital Speech Processing 9 a speaker and a listener who are in a common space.
From this point. This discretetime representation is the starting point for most applications. which can be as low as twice the cutoff frequency of the sharpcutoff digital filter. It is not an oversimplification to assert that digital speech processing is grounded in a set of techniques that have the goal of pushing the data rate of the speech representation to the left along either the upper or lower path in Figure 1. it is possible to incorporate aspects of both the speech production and speech perception process into the digital representation and processing. As we will see. these alternative representations are based on incorporating knowledge about the workings of the speech chain as depicted in Figure 1. and then reducing the sampling rate to the desired sampling rate.2 Applications of Digital Speech Processing The first step in most applications of digital speech processing is to convert the acoustic waveform to a sequence of numbers. 1. More interestingly for our purpose here — it is in this domain that we find the applications of digital speech processing. Most modern AtoD converters operate by sampling at a very high rate.2. applying a digital lowpass filter with cutoff set to preserve a prescribed bandwidth.2. other representations are obtained by digital processing. For the most part.make speech and message understanding more difficult in real communication environments. The remainder of this chapter is devoted to a brief summary of the .
e.applications of digital speech processing.” Figure 1. the systems that people interact with daily.3 Speech coding block diagram — encoder and decoder.3 shows a block diagram of a generic speech encoding/ . 1. 1.2. i. of speech signals.1 Speech Coding Perhaps the most widespread applications of digital speech processing technology occur in the areas of digital transmission and storage 10 Introduction AtoD Converter Analysis/ Encoding Channel or Medium Synthesis/ Decoding DtoA Converter speech signal samples data samples data xc (t) x[n] y[n] xc (t) x[n] y[n] decoded signal Fig.. It is common to refer to this activity as “speech coding” or “speech compression. In these areas the centrality of the digital representation is obvious. Our discussion will confirm the importance of the digital representation in all application areas. since the goal is to compress the digital waveform representation of speech into a lower bitrate representation.
This is the quintessential feature of digital coding. In the upper part of the figure. Because the digital representation at this point is often not directly related to the sampled speech waveform. The received data signal ˆy[n] is decoded using the inverse of the analysis processing. The digital signal x[n] is analyzed and coded by digital computation algorithms to produce a new digital signal y[n] that can be transmitted over a digital communication channel or stored in a digital storage medium as ˆy[n]. the transmitted (y[n]) and received (ˆy[n]) data can be essentially identical. As we will see. the AtoD converter converts the analog speech signal xc(t) to a sampled waveform representation x[n]. The lower path in Figure 1.2 Applications of Digital Speech Processing 11 With carefully designed error protection coding of the digital representation. there are a myriad of ways to do the encoding so as to reduce the data rate over that of the sampled and quantized speech waveform x[n].3 shows the decoder associated with the speech coder. .decoding (or compression) system. giving the sequence of samples ˆx[n] which is then converted (using a DtoA Converter) back to an analog signal ˆxc(t) for human listening. 1. y[n] and ˆy[n] are appropriately referred to as data signals that represent the speech signal. The decoder is often called a synthesizer because it must reconstitute the speech waveform from data that may bear no direct relationship to the waveform.
and prerecorded messages. the goal of the speech coder is to start with samples of the speech signal and reduce (compress) the data rate required to represent the speech signal while maintaining a desired perceptual fidelity.In theory. and for storage of speech for telephone answering machines. extremely narrowband communications channels (such as battlefield applications using high frequency (HF) radio). This means that the speech signal can be reconstructed to within the accuracy of the original coding for as long as the digital representation is retained. Speech coders enable a broad range of applications including narrowband and broadband wired telephony. In either case. perfect transmission of the coded digital representation is possible even under very noisy channel conditions. Speech coders often utilize many aspects of both . voice over internet protocol (VoIP) (which utilizes the internet as a realtime communications medium). or the bits saved can be devoted to error protection. cellular communications. interactive voice response (IVR) systems. and in the case of digital storage. secure voice for privacy and encryption (for national security applications). it is possible to store a perfect copy of the digital representation in perpetuity if sufficient care is taken to update the storage medium as storage technology advances. The compressed representation can be more efficiently transmitted or stored.
has the job of converting the printed text input . text and produce speech automatically. These coders are widely deployed in MP3 and AAC players and for audio in digital television systems [120].the speech production and speech perception processes.2 TexttoSpeech Synthesis For many years. The first block in the texttospeech synthesis system. a texttospeech synthesizer such as depicted in Figure 1. and hence may not be useful for more general audio signals such as music. In a sense.4 is a digital simulation of the entire upper part of the speech chain diagram.2.4 Texttospeech synthesis system block diagram. scientists and engineers have studied the speech production process with the goal of building a system that can start with 12 Introduction Linguistic Rules Synthesis Algorithm DtoA Converter text speech Fig. 1. labeled linguistic rules. Coders that are based on incorporating only aspects of sound perception generally do not achieve as much compression as those based on speech production. but they are more general and can be used for all types of audio signals. 1. The input to the system is ordinary text such as an email message or an article from a newspaper or magazine.
” In this method. pauses. the computer . how to pronounce ambiguous words like read. rates of speaking. the synthesis algorithm must simulate the action of the vocal tract system in creating the sounds of speech. specialized terms. The conversion from text to sounds involves a set of linguistic rules that must determine the appropriate set of sounds (perhaps including things like emphasis.) so that the resulting synthetic speech will express the words and intent of the text message in what passes for a natural voice that can be decoded accurately by human speech perception. There are many procedures for assembling the speech sounds and compiling them into a proper sentence. In essence. and how to properly pronounce proper names. Once the proper pronunciation of the text has been determined. This is more difficult than simply looking up the words in a pronouncing dictionary because the linguistic rules must determine how to pronounce acronyms. object. etc. (street or Saint). how to pronounce abbreviations like St. (Doctor or drive).into a set of sounds that the machine must synthesize. Dr. bass. but the most promising one today is called “unit selection and concatenation. etc. the role of the synthesis algorithm is to create the appropriate sound sequence to represent the text message in the form of speech.
The basic digital representation is not generally the sampled speech wave. etc. provide the voices for talking agents for completion of transactions over the internet. more importantly. dictionaries. to allow convenient manipulation of durations and blending of adjacent sounds. some sort of compressed representation is normally used to 1. serve as the voice for providing information from handheld devices such as foreign language phrasebooks. Instead. the speech synthesis algorithm would include an appropriate decoder. and then decides which sequence of speech units sounds best for the particular text message that is being produced. as discussed in Section 1. airline schedules. crossword puzzle helpers. syllables. etc.2 Applications of Digital Speech Processing 13 save memory and.).1. provide voice output from GPS systems in automobiles. half phones. Thus. handle call center help desks and customer care applications. Texttospeech synthesis systems are an essential component of modern human–machine communications systems and are used to do things like read email messages over a telephone. updates on arrivals and departures of flights. and as the voice of announcement machines that provide information such as stock quotes. whose output is converted to an analog representation via the DtoA converter.stores multiple versions of each of the basic units of speech (phones.2. .
5 shows a block diagram of a generic approach to pattern matching problems in speech processing.Another important application is in reading machines for the blind. where the object is to extract the message from the speech signal.5 Block diagram of general pattern matching system for speech signals. Such problems include the following: speech recognition. The first block in the pattern matching system converts the analog . speaker recognition. 1. 1. 14 Introduction signal.3 Speech Recognition and Other Pattern Matching Problems Another large class of digital speech processing applications is concerned with the automatic extraction of information from the speech signal. Figure 1. speaker verification. where an optical character recognition system provides the text input to a speech synthesis system. where the goal is to verify a speaker’s claimed identity from analysis of their speech AtoD Converter Feature Analysis Pattern Matching speech symbols Fig.2. which involves monitoring a speech signal for the occurrence of specified words or phrases. word spotting. Most such systems involve some sort of pattern matching. where the goal is to identify who is speaking. and automatic indexing of speech recordings based on recognition (or spotting) of spoken keywords.
Often. or a decision as to whether to accept or reject the identity claim of a speaker in the case of speaker verification. dynamically time aligns the set of feature vectors representing the speech signal with a concatenated set of stored patterns. and chooses the identity associated with the pattern which is the closest match to the timealigned set of feature vectors of the speech signal. The major areas where such a system finds applications include command and control of computer software. The feature analysis module converts the sampled speech signal to a set of feature vectors. voice dictation to create letters. The final block in the system. namely the pattern matching block. natural language voice dialogues with machines to enable help desks and . Although the block diagram of Figure 1.speech waveform to digital form using an AtoD converter. the biggest use has been in the area of recognition and understanding of speech in support of human– machine communication by voice. and other documents. the same analysis techniques that are used in speech coding are also used to derive the feature vectors. The symbolic output consists of a set of recognized words. or the identity of the best matching talker. in the case of speech recognition. memos. in the case of speaker recognition.5 represents a wide range of speech pattern matching problems.
etc. The goal of language translation systems is to convert spoken words in one language to spoken words in another language so as to facilitate natural language voice dialogues between people speaking different languages. Names from directories with upwards of several hundred names can readily be recognized and dialed using simple speech recognition technology. 1. Language translation technology requires speech synthesis systems that work in both languages. For example. Speech coding at bit rates on the order of 8 Kbps enables normal voice conversations in cell phones. along with speech recognition (and generally natural language understanding) that also works for both languages. one of the preeminent uses of speech technology is in portable communication devices.call centers. and for agent services such as calendar entry and update.2 Applications of Digital Speech Processing 15 Another major speech application that has long been a dream of speech researchers is automatic language translation. hence it is a very difficult task and one for which only limited progress . address list modification and entry. Pattern recognition applications often occur in conjunction with other digital speech processing applications. Spoken name speech recognition in cellphones enables voice dialing capability that can automatically dial the number associated with the recognized name.
4 Other Speech Applications The range of speech communication applications is illustrated in Figure 1. When such systems exist. As seen in this figure.7 represents any system where time signals such as speech are processed by the techniques of DSP. it will be possible for people speaking different languages to communicate at data rates on the order of that of printed text reading! 1. speech signal quality enhancement.6. it can be manipulated in virtually limitless ways by DSP techniques. and speech recognition as well as many others such as speaker identification.2.has been made. speech synthesis. Here again.or visuallyimpaired. This figure simply depicts the notion that once the speech signal is sampled. the techniques of digital speech processing are a key ingredient of a wide range of applications that include the three areas of transmission/storage. The block diagram in Figure 1. and aids for the hearing. manipulations and modifications of the speech signal are Digital Transmission & Storage Digital Speech Processing Techniques Speech Synthesis Speech Recognition Speaker Verification/ .
the goal of speech enhancement systems is to make the speech more intelligible and more natural. 1. In humantohuman communication. where the goal is to remove or suppress noise or echo or reverberation picked up by a microphone along with the desired speech signal. in reality the best .7 General block diagram for application of digital signal processing to speech signals. 1.Identification Enhancement of Speech Quality Aids for the Handicapped Fig. however. 16 Introduction AtoD Converter Computer Algorithm DtoA Converter speech speech Fig. One important application area is speech enhancement. operating on that representation by further digital computation.6 Range of speech communication applications. and then transforming back to the waveform domain. using a DtoA converter. usually achieved by transforming the speech signal into an alternative representation (that is motivated by our understanding of speech production and speech perception).
Success has been achieved. synthesizer. and we have hinted at some of the possibilities that exist for the future. We have given a brief overview of the way in which digital speech processing is being applied today. the intelligibility of the noisy speech. which we will discuss in the remainder of this text. but does not improve. rapid review of voice mail messages. These and many more examples all rely on the basic principles of digital speech processing. and to speedup or slowdown prerecorded speech (e. or careful scrutinizing of spoken material). or recognizer.3 Our Goal for this Text 17 too deep. in making distorted speech signals more useful for further processing as part of a speech coder. An excellent reference in this area is the recent textbook by Loizou [72]. to modify voice qualities.g. Other examples of manipulation of the speech signal include timescale modification to align voices with video segments. however. 1.that has been achieved so far is less perceptually annoying speech that essentially maintains. for talking books. Our goal is only to provide an uptodate introduction to this . We make no pretense of exhaustive coverage. The subject is too broad and 1..3 Our Goal for this Text We have discussed the speech signal and how it encodes information for human communication.
fascinating field. We will not be able to go into great depth, and we will not be able to cover all the possible applications of digital speech processing techniques. Instead our focus is on the fundamentals of digital speech processing and their application to coding, synthesis, and recognition. This means that some of the latest algorithmic innovations and applications will not be discussed — not because they are not interesting, but simply because there are so many fundamental triedandtrue techniques that remain at the core of digital speech processing. We hope that this text will stimulate readers to investigate the subject in greater depth using the extensive set of references provided. 2 The Speech Signal As the discussion in Chapter 1 shows, the goal in many applications of digital speech processing techniques is to move the digital representation of the speech signal from the waveform samples back up the speech chain toward the message. To gain a better idea of what this means, this chapter provides a brief overview of the phonetic representation of speech and an introduction to models for the production of the speech signal. 2.1 Phonetic Representation of Speech Speech can be represented phonetically by a finite set of symbols called the phonemes of the language, the number of which depends upon the
language and the refinement of the analysis. For most languages the number of phonemes is between 32 and 64. A condensed inventory of the sounds of speech in the English language is given in Table 2.1, where the phonemes are denoted by a set of ASCII symbols called the ARPAbet. Table 2.1 also includes some simple examples of ARPAbet transcriptions of words containing each of the phonemes of English. Additional phonemes can be added to Table 2.1 to account for allophonic variations and events such as glottal stops and pauses. 18 2.1 Phonetic Representation of Speech 19 Table 2.1 Condensed list of ARPAbet phonetic symbols for North American English. Class ARPAbet Example Transcription Vowels and IY beet [B IY T] diphthongs IH bit [B IH T] EY bait [B EY T] EH bet [B EH T] AE bat [B AE T] AA bob [B AA B] AO born [B AO R N] UH book [B UH K] OW boat [B OW T] UW boot [B UW T] AH but [B AH T ] ER bird [B ER D] AY buy [B AY] AW down [D AW N] OY boy [B OY] Glides Y you [Y UH]
R rent [R EH N T] Liquids W wit [W IH T] L let [L EH T] Nasals M met [M EH T] N net [N EH T] NG sing [S IH NG] Stops P pat [P AE T] B bet [B EH T] T ten [T EH N] D debt [D EH T] K kit [K IH T] G get [G EH T] Fricatives HH hat [HH AE T] F f at [F AE T] V vat [V AE T] TH thing [TH IH NG] DH that [DH AE T] S sat [S AE T] Z zoo [Z UW] SH shut [SH AH T] ZH az ure [AE ZH ER] Affricates CH chase [CH EY S] JH j udge [JH AH JH ] aThis set of 39 phonemes is used in the CMU Pronouncing Dictionary available online at http://www.speech.cs.cmu.edu/cgibin/cmudict. Figure 1.1 on p. 4 shows how the sounds corresponding to the text “should we chase” are encoded into a speech waveform. We see that, for the most part, phonemes have a distinctive appearance in the speech waveform. Thus sounds like /SH/ and /S/ look like (spectrally shaped) 20 The Speech Signal
random noise, while the vowel sounds /UH/, /IY/, and /EY/ are highly structured and quasiperiodic. These differences result from the distinctively different ways that these sounds are produced. 2.2 Models for Speech Production A schematic longitudinal crosssectional drawing of the human vocal tract mechanism is given in Figure 2.1 [35]. This diagram highlights the essential physical features of human anatomy that enter into the final stages of the speech production process. It shows the vocal tract as a tube of nonuniform crosssectional area that is bounded at one end by the vocal cords and at the other by the mouth opening. This tube serves as an acoustic transmission system for sounds generated inside the vocal tract. For creating nasal sounds like /M/, /N/, or /NG/, a sidebranch tube, called the nasal tract, is connected to the main acoustic branch by the trapdoor action of the velum. This branch path radiates sound at the nostrils. The shape (variation of crosssection along the axis) of the vocal tract varies with time due to motions of the lips, jaw, tongue, and velum. Although the actual human vocal tract is not laid out along a straight line as in Figure 2.1, this type of model is a reasonable approximation for wavelengths of the sounds in speech. The sounds of speech are generated in the system of Figure 2.1 in several ways. Voiced sounds (vowels, liquids, glides, nasals in Table 2.1) Fig. 2.1 Schematic model of the vocal tract system. (After Flanagan et al. [35].)
2.2 Models for Speech Production 21 are produced when the vocal tract tube is excited by pulses of air pressure resulting from quasiperiodic opening and closing of the glottal orifice (opening between the vocal cords). Examples in Figure 1.1 are the vowels /UH/, /IY/, and /EY/, and the liquid consonant /W/. Unvoiced sounds are produced by creating a constriction somewhere in the vocal tract tube and forcing air through that constriction, thereby creating turbulent air flow, which acts as a random noise excitation of the vocal tract tube. Examples are the unvoiced fricative sounds such as /SH/ and /S/. A third sound production mechanism is when the vocal tract is partially closed off causing turbulent flow due to the constriction, at the same time allowing quasiperiodic flow due to vocal cord vibrations. Sounds produced in this manner include the voiced fricatives /V/, /DH/, /Z/, and /ZH/. Finally, plosive sounds such as /P/, /T/, and /K/ and affricates such as /CH/ are formed by momentarily closing off air flow, allowing pressure to build up behind the closure, and then abruptly releasing the pressure. All these excitation sources create a wideband excitation signal to the vocal tract tube, which acts as an acoustic transmission line with certain vocal tract shapedependent resonances that tend to emphasize some frequencies of the excitation relative to others. As discussed in Chapter 1 and illustrated by the waveform in
1 can be described by acoustic theory.1. The system of Figure 2. the general character of the speech signal varies at the phoneme rate. the changes in vocal tract configuration occur relatively slowly compared to the detailed time variation of the speech signal. 34]. These resonance frequencies are called the formant frequencies of the sound [32. the fine structure of the time waveform is created by the sound sources in the vocal tract. and the resonances of the vocal tract tube shape these sound sources into the phonemes. . which is on the order of 10 phonemes per second. while the detailed time variations of the speech waveform are at a much higher rate. 2.2 Source/system model for a speech signal. and numerical techniques can be used to create a complete physical 22 The Speech Signal Linear System Excitation Generator excitation signal speech signal Vocal Tract Parameters Excitation Parameters Fig.Figure 1. In summary. That is. The resonance frequencies resulting from a particular configuration of the articulators are instrumental in forming the sound corresponding to a given phoneme. The sounds created in the vocal tract are shaped in the frequency domain by the frequency response of the vocal tract.
In general such a model is called a source/system model of speech production. and since the vocal tract changes shape relatively slowly. Thus. but. 93]. it is sufficient to model the production of a sampled speech signal by a discretetime system model such as the one depicted in Figure 2. The discretetime timevarying linear system on the right in Figure 2. The excitation generator on the left simulates the different modes of sound generation in the vocal tract. The shorttime frequency response of the linear system simulates the frequency shaping of the vocal tract system. Samples of a speech signal are assumed to be the output of the timevarying linear system.2 simulates the frequency shaping of the vocal tract tube. it is common to characterize the discretetime linear system by a system function of the form: H(z) = M_ k=0 bkz −k 1− _N k=1 akz . it is reasonable to assume that the linear system response does not vary over time intervals on the order of 10 ms or so. for the most part.2.simulation of sound generation and transmission in the vocal tract [36.
−k = b0 M_ k=1 (1 − dkz −1) N_ k=1 (1 − ckz −1) . Some of the poles (ck) of the system function lie close to the unit circle and create resonances to model the formant frequencies.1) where the filter coefficients ak and bk (labeled as vocal tract parameters in Figure 2. as we discuss further in Chapter 4. The box labeled excitation generator in Figure 2. However.2 creates an appropriate excitation for the type of sound being produced. many applications of the source/system model only include poles in the model because this simplifies the analysis required to estimate the parameters of the model from the speech signal. it is sometimes useful to employ zeros (dk) of the system function to model nasal and fricative 2. 64]. In detailed modeling of speech production [32.2 Models for Speech Production 23 sounds.2) change at a rate on the order of 50–100 times/s. 34. For voiced speech the excitation to the linear system is a quasiperiodic sequence . (2.
2 switches from unvoiced to voiced leading to the speech signal output as shown in the figure. and a wide variety of digital representations of the speech . the linear system imposes its frequency response on the spectrum to create the speech sounds. This model of speech as the output of a slowly timevarying digital filter with an excitation that captures the nature of the voiced/unvoiced distinction in speech production is the basis for thinking about the speech signal. Often it is convenient to absorb the glottal pulse spectrum contribution into the vocal tract system model of (2. The individual finiteduration glottal pulses have a lowpass spectrum that depends on a number of factors [105].1). For unvoiced speech.of discrete (glottal) pulses that look very much like those shown in the righthand half of the excitation signal waveform in Figure 2. This can be achieved by a small increase in the order of the denominator over what would be needed to represent the formant resonances. In either case. the linear system is excited by a random number generator that produces a discretetime noise signal with flat spectrum as shown in the lefthand half of the excitation signal. The excitation in Figure 2. Therefore. The fundamental frequency of the glottal excitation determines the perceived pitch of the voice.2. the periodic sequence of smooth glottal pulses has a harmonic line spectrum with components that decrease in amplitude with increasing frequency.
signal are based on it. such models are based on many approximations including the assumption that the source and the system do not interact. That is.2 with the system characterized by a timesequence of timeinvariant systems are quite sufficient for most applications in speech processing. It is through such models and analysis techniques that we are able to build properties of the speech production process into digital representations of the speech signal. By assuming that the properties of the speech signal (and the model) are constant over short time intervals. the assumption of linearity. it is possible to compute/measure/estimate the parameters of the model by analyzing short blocks of samples of the speech signal.3 More Refined Models Source/system models as shown in Figure 2. However. Fluid mechanics and acoustic wave propagation theory are fundamental physical principles that must be applied for detailed modeling of speech production. the speech signal is represented by the parameters of the model instead of the sampled waveform. Since the early work of Flanagan and Ishizaka [34. 36. and we shall rely on such models throughout this text. 51] . 24 The Speech Signal 2. and the assumption that the distributed continuoustime vocal tract transmission system can be modeled by a discrete linear timeinvariant system.
In this chapter we turn to the perception side of the speech chain to discuss properties of human sound perception that can be employed to create digital representations of the speech signal that are perceptually robust. Recent advances in MRI imaging and computer image analysis have provided significant advances in this area of speech science [17]. Early efforts to measure vocal tract area functions involved hand tracing on Xray pictures [32].1 shows a schematic view of the human ear showing the three . For many years. Stevens [121] and Quatieri [94] provide useful discussions of these effects. 3. we introduced the speech production process and showed how we could model speech production using discretetime systems.much work has been devoted to creating detailed simulations of glottal flow. researchers have sought to measure the physical dimensions of the human vocal tract during speech production. 3 Hearing and Auditory Perception In Chapter 2. and the nonlinearities that enter into sound generation and transmission in the vocal tract.1 The Human Ear Figure 3. This information is essential for detailed simulations based on acoustic theory. the interaction of the glottal source and the vocal tract in speech production.
and finally.distinct sound processing sections. the inner ear. 3. the malleus (also called the hammer).2 [107] depicts a block diagram abstraction of the auditory system. Distributed along Fig. which consists of the cochlea and the set of neural connections to the auditory nerve. the middle ear beginning at the tympanic membrane. (After Flanagan [34]. the incus (also called the anvil) and the stapes (also called the stirrup). or eardrum. which gathers sound and conducts it through the external canal to the middle ear. 3.1 Schematic view of the human ear (inner and middle structures enlarged). which perform a transduction from acoustic waves to mechanical pressure waves. The basilar membrane vibrates in a frequencyselective manner along its extent and thereby performs a rough (nonuniform) spectral analysis of the sound. which conducts the neural signals to the brain. (After Sachs et al. and including three small bones. The acoustic wave is transmitted from the outer ear to the inner ear where the ear drum and bone structures convert the sound wave to mechanical vibrations which ultimately are transferred to the basilar membrane inside the cochlea.2 Schematic model of the auditory mechanism.) . namely: the outer ear consisting of the pinna. 25 26 Hearing and Auditory Perception Fig. [107].) Figure 3.
The processing at higher levels in the brain.2 Perception of Loudness A key factor in the perception of speech and other sounds is loudness. Loudness is a perceptual quality that is related to the physical property of sound pressure level. This produces an auditory nerve representation in both time and frequency.3. This relationship . 3. These experiments have yielded much valuable knowledge about the sensitivity of the human auditory system to acoustic properties such as intensity and frequency.2 Perception of Loudness 27 the basilar membrane are a set of inner hair cells that serve to convert motion along the basilar membrane to neural activity. Even so.2 as a sequence of central processing with multiple representations followed by some type of pattern recognition. is not well understood and we can only postulate the mechanisms used by the human brain to perceive sound or speech. shown in Figure 3. Loudness is quantified by relating the actual sound pressure level of a pure tone (in dB relative to a standard reference level) to the perceived loudness of the same tone (in a unit called phons) over the range of human hearing (20 Hz–20 kHz). a wealth of knowledge about how sounds are perceived has been discovered by careful experiments that use tones and noise signals to stimulate the auditory system of human observers in very specific and controlled ways.
(After Fletcher and Munson [37].3 Loudness level for human hearing. Specifically. both the 50 dB 1000 Hz tone and the 60 dB 100 Hz tone are said to have a loudness level of 50 phons . the point at frequency 100 Hz on the curve labeled 50 (phons) is obtained by adjusting the power of the 100 Hz tone until it sounds as loud as a 1000 Hz tone having a sound pressure level of 50 dB. The solid curves are equalloudnesslevel contours measured by comparing sounds at various frequencies with a pure tone of frequency 1000 Hz and known sound pressure level. the dotted curve at the bottom of the figure labeled “threshold of audibility” shows the sound pressure level that is required for a sound of a given frequency to be just audible (by a person with normal hearing). It can be seen that low frequencies must be significantly more intense than frequencies in the midrange in order that they be perceived at all.is shown in Figure 3. 103].) to be perceived to be equal in loudness to the 1000 Hz tone of sound pressure level 50 dB. By convention. 3. For example.3 [37. Careful measurements of this kind show that a 100 Hz tone must have a sound pressure level of about 60 dB in order 28 Hearing and Auditory Perception Fig. These loudness curves show that the perception of loudness is frequencydependent.
4. 3.4. 3. and with a relative bandwidth of about 20% of the center frequency above 500 Hz. This is almost precisely the range of frequencies occupied by most of the sounds of speech. Even so. In reality. the bandpass filters are not ideal 3. An idealized version of such a filter bank is depicted schematically in Figure 3. The equalloudnesslevel curves show that the auditory system is most sensitive for frequencies ranging from about 100 Hz up to about 6 kHz with the greatest sensitivity at around 3 to 4 kHz.4 Schematic representation of bandpass filters according to the critical band theory of hearing. but their frequency responses overlap significantly since points on the basilar membrane cannot vibrate independently of each other. showing that the effective bandwidths are constant at about 100 Hz for center frequencies below 500 Hz.(pronounced as /F OW N Z/). and the critical bandwidths have been defined and measured using a variety of methods. the concept of bandpass filter analysis in the cochlea is well established.3 Critical Bands The nonuniform frequency analysis performed by the basilar membrane can be thought of as equivalent to that of a set of bandpass filters whose frequency responses become increasingly broad with increasing frequency. An equation that fits empirical .4 Pitch Perception 29 amplitude Fig. as shown in Figure 3.
pitch perception. Approximately 25 critical band filters span the range from 0 to 20 kHz. Like loudness.5 Relation between subjective pitch and frequency of a pure tone.measurements over the auditory range is Δfc =25 + 75[1 + 1.1) where Δfc is the critical bandwidth associated with center frequency fc [134]. 3.2) which is plotted in Figure 3. and it therefore provides motivation for digital representations of the speech signal that are based on a frequency decomposition. The relationship between pitch (measured 30 Hearing and Auditory Perception Fig. The concept of critical bands is very important in understanding such phenomena as loudness perception. (3. This expression is calibrated so that a frequency of 1000 Hz corresponds to a pitch of 1000 mels. on a nonlinear frequency scale called the melscale) and frequency of a pure tone is approximated by the equation [122]: Pitch in mels = 1127loge(1 + f/700). 3. pitch is a subjective attribute of sound that is related to the fundamental frequency of the sound. and such sounds are perceived by the auditory system as having a quality known as pitch. This empirical . and masking.5. (3.69.4 Pitch Perception Most musical sounds as well as voiced speech sounds have a periodic structure when viewed over short time intervals. which is a physical attribute of the acoustic waveform [122].4(fc/1000)2]0.
Thus. where a critical band of width Δfc = 160 Hz centered on fc = 1000 Hz maps into a band of width 106 mels and a critical band of width 100 Hz centered on 350 Hz maps into a band of width 107 mels. 3. Below 1000 Hz. This is shown in Figure 3. For example. the relationship between pitch and frequency is nearly proportional. what we know about pitch perception reinforces the notion that the auditory system performs a frequency analysis that can be simulated with a bank of bandpass filters whose bandwidths increase as center frequency increases. a tone of frequency 1000 Hz is given a pitch of 1000 mels. can be related to the concept of critical bands [134]. but contains many frequencies.5 Auditory Masking 31 Voiced speech is quasiperiodic. however. To calibrate the scale. the relationship is nonlinear. Nevertheless. one critical bandwidth corresponds to about 100 mels on the pitch scale.scale describes the results of experiments where subjects were asked to adjust the pitch of a measurement tone to half the pitch of a reference tone. For higher frequencies. It turns out that more or less independently of the center frequency of the band. as quantified by the melscale. The psychophysical phenomenon of pitch.2) shows that a frequency of f = 5000 Hz corresponds to a pitch of 2364 mels.5. many of the results obtained with pure tones are relevant to . (3.
Pure tones can mask other pure tones. As shown in Figure 3. 3. Figure 3. and noise can mask pure tones as well.5 Auditory Masking The phenomenon of critical band auditory analysis can be explained intuitively in terms of vibrations of the basilar membrane. A detailed discussion of masking can be found in [134]. Often the term pitch period is used for the fundamental period of the voiced speech signal even though its usage in this way is somewhat imprecise. is also attributable to the mechanical vibrations of the basilar membrane. called masking.the perception of voice pitch as well.6. All spectral components whose level is below this raised threshold are masked and therefore do Masker Shifted threshold Threshold of hearing . The notion that a sound becomes inaudible can be quantified with respect to the threshold of audibility. Masking occurs when one sound makes a second superimposed sound inaudible. an intense tone (called the masker) tends to raise the threshold of audibility around its location on the frequency axis as shown by the solid line. Loud tones causing strong vibrations at a point on the basilar membrane can swamp out vibrations that occur nearby. A related phenomenon.6 illustrates masking of tones by tones.
6 Illustration of effects of masking. 3. This is shown in Figure 3. it is possible to achieve lower data rate representations while maintaining a high degree of perceptual fidelity. 32 Hearing and Auditory Perception not need to be reproduced in a speech (or audio) processing system because they would not be heard. Similarly.6 Complete Model of Auditory Processing In Chapter 2. and therefore will be heard. any spectral component whose level is above the raised threshold is not masked. we described elements of a generative model of speech . Masking is widely employed in digital representations of speech (and audio) signals by “hiding” errors in the representation in areas where the threshold of hearing is elevated by strong frequency components in the signal [120].Masked signals (dashed) Unmasked signal log frequency amplitude Fig. 3.6 where the falloff of the shifted threshold is less abrupt above than below the masker. In this way. It has been shown that the masking effect is greater for frequencies above the masking frequency than below.
2. and thus we rely on psychophysical experimentation to understand the role of loudness. 115] and used in a range of speech processing systems. pitch perception. we presented a model for speech production in which an excitation source provides the basic temporal fine structure while a slowly varying filter provides spectral shaping (often referred to as the spectrum envelope) to create the various sounds of speech. and auditory masking in speech perception in humans. However. 48. is rudimentary at best. we described elements of a model of speech perception. Both the excitation and the linear system were defined implicitly by the assertion that the sampled speech signal was the output . and few details of the excitation or the linear system were given. beyond the basilar membrane processing of the inner ear. completely describe and model the ways in which speech is produced by humans. critical bands. the source/system separation was presented at an abstract level. In Figure 2. 4 ShortTime Analysis of Speech In Figure 2. in theory. 73.production which could.2 of Chapter 2. Although some excellent auditory models have been proposed [41. all such models are incomplete representations of our knowledge about how speech is understood. In this chapter. the problem that arises is that our detailed knowledge of how speech is perceived and understood.
Clearly. and the voiced excitation is assumed to be a periodic impulse train with impulses spaced by the pitch period (P0) rounded to the nearest 33 34 ShortTime Analysis of Speech Fig. sample. In this model the unvoiced excitation is assumed to be a random noise sequence.1 The pulses needed to model the glottal flow waveform during voiced speech are assumed to be combined (by convolution) with the impulse response of the linear system.1 Voiced/unvoiced/system model for a speech signal.1. the impulse response. Since our goal is to extract parameters of the model by analysis of the speech signal. One such model uses a more detailed representation of the excitation in terms of separate source generators for voiced and unvoiced speech as shown in Figure 4. it is common to assume structures (or representations) for both the excitation generator and the linear system. which is assumed to be slowlytimevarying (changing every 50–100 ms or so). For example over time intervals of tens of milliseconds.of the overall system. this is not sufficient to uniquely specify either the excitation or the system. frequency response.1) . (4. 4. the system can be described by the convolution expression sˆn[n] = ∞_ m=0 hˆn[m]eˆn[n − m]. By this we mean that over the timescale of phonemes. and system function of the system remains relatively constant.
it is often assumed that the system is an allpole system with 1 As mentioned in Chapter 3. (4. However. the most general linear timeinvariant system would be characterized by a rational system function as given in (2. to simplify analysis.2) change with time. 35 system function of the form: H(z) = G 1− _p k=1 akz −k . the gain Gˆn is absorbed into hˆn[m] for convenience.1). but this complication of notation is not generally necessary since it is usually clear that the system function or difference equation only applies . and should therefore be indexed with the subscript ˆn as in (4. As discussed in Chapter 2.1) and Figure 4.where the subscript ˆn denotes the time index pointing to the block of samples of the entire speech signal s[n] wherein the impulse response hˆn[m] applies.We use n for the time index within that interval.2) The coefficients G and ak in (4. and m is the index of summation in the convolution sum.1. In this model. the period of voiced speech is related to the fundamental frequency (perceived as pitch) of the voice.
over a short time interval. it is common to process speech in blocks (also called “frames”) over which the properties of the speech waveform can be assumed to remain relatively constant. but whenever we want to . the input and output are related by a difference equation of the form: s[n] = _p k=1 aks[n − k] + Ge[n].3) where as discussed above. as represented by (4. Because of the slowly varying nature of the speech signal. and techniques for extracting the linear system model (which imparts the spectrum envelope or spectrum shaping). the linear system in the model is commonly referred to as simply the “vocal tract” system and the corresponding impulse response is called the “vocal tract impulse response. vocal tract tube. we have suppressed the indication of the time at which the difference equation applies.” For allpole linear systems. This leads to the basic principle of shorttime analysis. such as the pitch period and voiced/unvoiced classification.2 Although the linear system is assumed to model the composite spectrum effects of radiation. common practice is to partition the analysis of the speech signal into techniques for extracting the parameters of the excitation model. which 2 In general.2). we use n and m for discrete indices for sequences. and glottal excitation pulse shape (for voiced speech only) over a short time interval. With such a model as the basis. (4.
e.5) Figure 4. The infinite limits in (4. For example. and w[ˆn − m] represents a timeshifted window sequence. a finiteduration4 window might be a Hamming window defined by wH[m] = _ 0.46cos(πm/M) −M ≤ m ≤ M 0 otherwise. i. (4.indicate a specific analysis time.4) imply summation over all nonzero values of the windowed segment xˆn[m] = x[m]w[ˆn − m].4) where Xˆn represents the shorttime analysis parameter (or vector of parameters) at analysis time ˆn.We will see several examples of such operators that are designed to extract or highlight certain features of the speech signal.54 + 0. for all m in the region of support of the window.. we use ˆn. 36 ShortTime Analysis of Speech is represented in a general form by the equation: Xˆn = ∞_ m=−∞ T {x[m]w[ˆn − m]}. (4.3 The operator T{ } defines the nature of the shorttime analysis function.2 shows a discretetime Hamming window and its discretetime Fourier transform.5 It can be shown . whose purpose is to select a segment of the sequence x[m] in the neighborhood of sample m = ˆn. as used throughout this chapter.
Other windows will have similar properties. Figure 4.2 Hamming window (a) and its discretetime Fourier transform (b). 4An example of an infiniteduration window is w[n] = nan for n ≥ 0. Such a window can lead to recursive implementations of shorttime analysis functions [9].3 shows a 125 ms segment of a speech waveform that includes both unvoiced (0–50 ms) and voiced speech (50–125 ms). 0 20 40 60 80 100 120 time in ms .that a (2M + 1) sample Hamming window has a frequency main lobe (full) bandwidth of 4π/M. they will be concentrated in time and frequency. and the frequency width will be inversely proportional to the time width [89]. Also shown is a sequence of data windows of duration 40 ms and shifted by 15 ms (320 samples at 16 kHz sampling rate) between windows. 4.. A causal window would be shifted to the right by M samples. _ ∞ m=−∞w[m]..1 ShortTime Energy and ZeroCrossing Rate 37 (a) (b) Fig.e.e. i. This illustrates how shorttime analysis is implemented. 5We have assumed the time origin at the center of a symmetric interval of 2M + 1 samples. so that Xˆn is a weighted average. i.4) is normalized by dividing by the effective window length. 4. 3 Sometimes (4.
4. 4.3 Section of speech waveform with shorttime analysis windows. Eˆn = x2[n] ∗ he[n] __ n=ˆn. These functions are simple to compute. the shorttime zero crossing rate is defined as the weighted average of the number of times the speech signal changes sign within the time window. (4. In this case. it is often possible to express shorttime analysis operators as a convolution or linear filtering operation. and they are useful for estimating properties of the excitation function in the model. Similarly.1 ShortTime Energy and ZeroCrossing Rate Two basic shorttime analysis functions useful for speech signals are the shorttime energy and the shorttime zerocrossing rate. Representing this operator in terms of linear filtering leads to .Fig. As shown in (4.6) 38 ShortTime Analysis of Speech In this case the operator T{ } is simply squaring the windowed samples. The shorttime energy is defined as Eˆn = ∞_ m=−∞ (x[m]w[ˆn − m])2 = ∞_ m=−∞ x2[m]w2[ˆn − m]. where the impulse response of the linear filter is he[n] = w2[n].6).
both the shorttime energy and the shorttime zerocrossing rate are output of a lowpass filter whose frequency response is as shown in Figure 4.7) where sgn{x} = _ 1x≥0 −1 x < 0.2(b). the frequency response is very .5sgn{x[m]} − sgn{x[m − 1]}w[ˆn − m]. Figure 4. (4. While this is a convenient representation that fits the general framework of (4. the window is a Hamming window (two examples shown) of duration 25 ms (equivalent to 401 samples at a 16 kHz sampling rate). (4. the computation of Zˆn could be implemented in other ways. In both cases.5sgn{x[m]} − sgn{x[m − 1]} is equal to 1 if x[m] and x[m − 1] have different algebraic signs and 0 if they have the same sign.Zˆn = ∞_ m=−∞ 0.4 shows an example of the shorttime energy and zerocrossing rate for a segment of speech with a transition from unvoiced to voiced speech.4). For the 401point Hamming window used in Figure 4.6 Thus.7) is a weighted sum of all the instances of alternating sign (zerocrossing) that fall within the support region of the shifted window w[ˆn − m]. it follows that Zˆn in (4.8) Since 0.4.
the window applied to the signal samples was the squareroot of the Hamming window. Conversely. For finitelength windows like the Hamming window. Note that during the unvoiced interval. so that he[n] = w2[n] is the Hamming window defined by (4. the zerocrossing rate is relatively high compared to the zerocrossing rate in the voiced interval. This is . they can be sampled at a much lower rate than that of the original speech signal. the energy is relatively low in the unvoiced region compared to the energy in the voiced region.4 Section of speech waveform with shorttime energy and zerocrossing rate superimposed.1 ShortTime Energy and ZeroCrossing Rate 39 0 20 40 60 80 100 120 time in ms Fig. Note also that there is a small shift of the two curves relative to events in the time waveform. and therefore. 4. This means that the shorttime 6 In the case of the shorttime energy.5). this reduction of the sampling rate is accomplished by moving the window position ˆn in jumps of more than one sample as shown in Figure 4. energy and zerocrossing rate functions are slowly varying compared to the time variations of the speech signal. 4.3.small for discretetime frequencies above 2π/200 rad/s (equivalent to 16000/200 = 80Hz analog frequency).
due to the time delay of M samples (equivalent to 12.2 ShortTime Autocorrelation Function (STACF) The autocorrelation function is often used as a means of detecting periodicity in signals. 4. A complete algorithm would involve measurements of the statistical distributions of the energy and zerocrossing rate for both voiced and unvoiced speech segments (and also background noise distributions). and they are simple to compute. From our model. we expect unvoiced regions to have lower shorttime energy than voiced regions. The shorttime energy is an indication of the amplitude of the signal in the interval around time ˆn. The shorttime energy and shorttime zerocrossing rate are important because they abstract valuable information about the speech signal. the shorttime energy and shorttime zerocrossing rate can be the basis for an algorithm for making a decision as to whether the speech signal is voiced or unvoiced at 40 ShortTime Analysis of Speech any particular time ˆn. Voiced signals have a high frequency (HF) falloff due to the lowpass nature of the glottal pulses. and it is also the basis for many spectrum analysis . the shorttime zerocrossing rate is a crude frequency analyzer. Similarly. Thus. These distributions can be used to derive thresholds used in voiced/unvoiced decision [100].5 ms) added to make the analysis window filter causal. while unvoiced sounds have much more HF energy.
9]. (4.methods. Note that the STACF is a twodimensional function of the discretetime index ˆn (the window position) and the discretelag index _. (4. the shorttime autocorrelation of (4. For infiniteduration decaying exponential windows.9) can be expressed in terms of linear timeinvariant (LTI) filtering as φˆn[_] = ∞_ m=−∞ x[m]x[m − _] ˜ w_[ˆn − m]. If the window has finite duration.9) Using the familiar evensymmetric property of the autocorrelation.e.6). i. φˆn[_] = ∞_ m=−∞ xˆn[m]xˆn[m + _] = ∞_ m=−∞ x[m]w[ˆn − m]x[m + _]w[ˆn − m − _]. (4.. To see how the shorttime autocorrelation can be used in speech . φˆn[−_] = φˆn[_]. This makes it a useful tool for shorttime speech analysis.9) can be evaluated directly or using FFT techniques (see Section 4.10) can be computed recursively at time ˆn by using a different filter ˜ w_[m] for each lag value _ [8. (4. The STACF is defined as the deterministic autocorrelation function of the sequence xˆn[m] = x[m]w[ˆn − m] that is selected by the window shifted to time ˆn.10) where ˜ w_[m] = w[m]w[m + _].
assume that a segment of the sampled speech signal is a segment of the output of the discretetime model shown in Figure 4. In the case of the speech signal.. vocal tract shape. property of the autocorrelation function is that φ(s)[_] = φ(e)[_] ∗ φ(h)[_]. (4.2 ShortTime Autocorrelation Function (STACF) 41 the system is characterized at a particular analysis time by an impulse response h[n]. assume that s[n] = e[n] ∗ h[n]. That is.analysis.) Different segments of the speech signal will have the same form of model with different excitation and system impulse response. A well known. and the input is either a periodic impulse train or random white noise. the autocorrelation function of s[n] = e[n] ∗ h[n] is the convolution of the autocorrelation functions of e[n] and h[n]. and radiation at the lips. therefore. the autocorrelation of a periodic impulse train excitation with period P0 is a periodic impulse train sequence with the same period. (Note that we have suppressed the indication of analysis time. and easily proved. For voiced speech. where e[n] is the excitation to the linear system with impulse response h[n].e. the autocorrelation of voiced speech is the periodic autocorrelation function φ(s)[_] = ∞_ . In this case.1 where 4. h[n] represents the combined (by convolution) effects of the glottal pulse shape (for voiced speech).11) i.
13) Equation (4. (4.12) and (4. whose stochastic autocorrelation function would be an impulse sequence at _ = 0.2 0 10 20 30 40 −0.2 −0.1 .1 0 0. However.m=−∞ φ(h)[_ − mP0]. the excitation can be assumed to be random white noise.13) assumes probability average or averaging over an infinite time interval (for stationary random signals).1 0.2 −0.13) except that the correlation values will taper off with lag _ due to the tapering of the window and the fact that less and less data is involved in the computation of the shorttime autocorrelation for longer 42 ShortTime Analysis of Speech 0 10 20 30 40 −0. and (4. Therefore.12) In the case of unvoiced speech.1 0 0. the autocorrelation function of unvoiced speech computed using averaging would be simply φ(s)[_] = φ(h)[_]. (4. the deterministic autocorrelation function of a finitelength segment of the speech waveform will have properties similar to those of (4.12) assumes periodic computation of an infinite periodic signal.
lag values. Note the peak in the autocorrelation function for the voiced segment at the pitch period and twice the pitch period. Finally.5 1 (a) Voiced Segment (b) Voiced Autocorrelation (c) Unvoiced Segment (d) Unvoiced Autocorrelation time in ms time in ms Fig.0.5 1 0 20 40 −0. Usually such algorithms involve the autocorrelation function and other shorttime measurements such as zerocrossings and energy to aid in making the voiced/unvoiced decision.2 0 10 20 30 40 −0.5 0 0.5 0 0.5 for both a voiced and an unvoiced speech segment.5 Voiced and unvoiced segments of speech and their corresponding STACF. 4. observe that the STACF implicitly contains the shorttime energy since Eˆn = ∞_ . and note the absence of such peaks in the autocorrelation function for the unvoiced segment. This suggests that the STACF could be the basis for an algorithm for estimating/detecting the pitch period of speech. This tapering off in level is depicted in Figure 4.
15) is a sequence of DTFTs. Thus. the STFT can be expressed in terms of a linear filtering operation. the STFT is a function of two variables. we will find that the STFT defined as Xˆn(ej ˆω) = ∞_ m=−∞ x[m]w[ˆn − m]e −j ˆωm. (4. i. for fixed analysis time ˆn.e. 89.15) can be expressed as the discrete ..m=−∞ (x[m]w[ˆn − m])2 = φˆn[0].15) is the basis for a wide range of speech analysis.7 Since (4. 129]. We now turn our attention to what is perhaps the most important basic concept in digital speech processing. In subsequent chapters. ˆn the discretetime index denoting the window position. the STFT is the discretetime Fourier transform (DTFT) of the signal xˆn[m] = x[m]w[ˆn − m].14) 4. (4.3 ShortTime Fourier Transform (STFT) The shorttime analysis functions discussed so far are examples of the general shorttime analysis principle that is the basis for most 4. For example. the DTFT of the (usually finiteduration) signal selected and amplitudeweighted by the sliding window w[ˆn − m] [24. the twodimensional function Xˆn(ej ˆω) at discretetime ˆn is a periodic function of continuous radian frequency ˆω with period 2π [89].3 ShortTime Fourier Transform (STFT) 43 algorithms for speech processing. and ˆω representing the analysis frequency. As in the case of the other shorttime analysis functions discussed in this chapter. coding and synthesis systems. By definition. (4.
and the window (lowpass) filter selects the resulting . (4.2(b). and ˆω is used to distinguish the STFT analysis frequency from the frequency variable of the nontimedependent Fourier transform frequency variable ω. when viewed as a linear filter impulse response. alternatively. Xˆn(ej ˆω) = _ x[n] ∗ (w[n]ej ˆωn) _ e −j ˆωn ___ n=ˆn . has a lowpass frequency response with the cutoff frequency varying inversely with the window length. (4. (See Figure 4.16) can be interpreted as follows: the amplitude modulation x[n]e−j ˆωn shifts the spectrum of 7 As before.convolution Xˆn(ej ˆω) = (x[n]e −j ˆωn) ∗ w[n] ___ n=ˆn . we use ˆn to specify the analysis time. Xˆn(ej ˆω) is slowly varying as ˆn varies. 44 ShortTime Analysis of Speech x[n] down by ˆω. Equation (4.16) or.17) Recall that a typical window like a Hamming window.) This means that for a fixed value of ˆω.
4. Then that band of frequencies is shifted down by the amplitude modulation with e−j ˆωn. The STFT becomes a practical tool for both analysis and applications when it is implemented with a finiteduration window moved in steps of R >1 samples in time and computed at a discrete set of frequencies as in XrR[k] = _rR . which selects the band of frequencies centered on ˆω. resulting again in the same lowpass signal [89]. (3) For each frequency ˆω. the STFT has three interpretations: (1) It is a sequence of discretetime Fourier transforms of windowed signal segments. (2) For each frequency ˆω with ˆn varying.4 Sampling the STFT in Time and Frequency As defined in (4.band of frequencies around zero frequency. In summary.. it is the time sequence output of a lowpass filter that follows frequency downshifting by ˆω. i. This is. the band of frequencies of x[n] that were originally centered on analysis frequency ˆω. An identical conclusion follows from (4.15). a periodic function of ˆω at each window position ˆn.17): x[n] is the input to a bandpass filter with impulse response w[n]ej ˆωn. of course.e. the STFT is a function of a continuous analysis frequency ˆω. it is the time sequence output resulting from frequency downshifting the output of a bandpass filter.
18) in the equivalent form: XrR[k] = ˜XrR[k]e −j(2πk/N)rR k = 0. ˜XrR[k]. and the segment of the speech signal is the timereversed sequence of L samples that precedes the analysis time. due to the definition of the window.18) where N is the number of uniformly spaced frequencies across the interval 0 ≤ ˆω < 2π. and L is the window length (in samples).. the analysis time rR is shifted to the time origin of the DFT computation. for specificity.. it is helpful to write (4.N − 1.20) Since we have assumed...N − 1..1.4 Sampling the STFT in Time and Frequency 45 where ˜X rR[k] = L_−1 m=0 x[rR − m]w[m]ej(2πk/N)m k = 0..1. has the interpretation of an Npoint DFT of the sequence x[rR − m]w[m]. that w[m] _= 0 only in the range 0 ≤ m ≤ L − 1..19) results from the shift of the time origin. the alternative form. . The complex exponential factor e−j(2πk/N)rR in (4... To aid in interpretation.N − 1. (4.8 In (4. Note that we have assumed that w[m] is causal and nonzero only in the range 0 ≤ m ≤ L − 1 so that the windowed segment x[m]w[rR − m] is nonzero over rR − L + 1 ≤ m ≤ rR..1.20). (4. is nonzero in the interval 0 ≤ m ≤ L − 1..m=rR−L+1 x[m]w[rR − m]e −j(2πk/N)m k = 0. (4.. which.19) 4.
(This can be done efficiently with an Npoint FFT algorithm. R.For a discretetime Fourier transform interpretation. (2) N ≥ L. (4) Move the time origin by R samples (i.. N. (4.e.. etc.L − 1.1.. 46 ShortTime Analysis of Speech following constraints on R and N: (1) R ≤ L/(2C) where C is a constant that is dependent on the window frequency bandwidth. and (3). Constraint (1) above is related to sampling the STFT in time at a rate of twice the window bandwidth in frequency in order to eliminate . C = 2 for a Hamming window.. It can easily be shown that both R and N are determined entirely by the time width and frequency bandwidth of the lowpass window.. (2) Compute the complex conjugate of the Npoint DFT of the sequence xrR[m]. The remaining issue for complete specification of the sampled STFT is specification of the temporal sampling period. and the number of uniform frequencies. r → r + 1) and repeat steps (1). used to compute the STFT [1].20) is the complex conjugate of the DFT of the windowed sequence [89]. C = 1 for a rectangular window. (2). w[m].) (3) The multiplication by e−j(2πk/N)rR can be done if necessary. but often can be omitted (as in computing a sound spectrogram or spectrographic display). where L is the window length in samples. Thus. since x[rR − m]w[m] is real. for m = 0. it follows that ˜X rR[k] can be computed by the following process: (1) Form the sequence xrR[m] = x[rR − m]w[m]. giving the 8 The DFT is normally defined with a negative exponent.
6 are simply a display of the magnitude of the STFT. Sound spectrograms like those in Figure 4. spectrograms were made by an ingenious device comprised of an audio tape loop. Such a function of two variables can be plotted on a two dimensional surface (such as this text) as . Today spectrograms like those in Figure 4.6.6 are made by DSP techniques [87] and displayed as either pseudocolor or grayscale images on computer screens.fk) = 20log10 ˜XrR[k] = 20log10 XrR[k]. N. which may be much larger than the window length L. Up until the 1970s. and constraint (2) above is related to sampling in frequency at a rate of twice the equivalent time width of the window to ensure that there is no aliasing in time of the STFT. the images in Figure 4.21) where the plot axes are labeled in terms of analog time and frequency through the relations tr = rRT and fk = k/(NT). variable analog bandpass filter. Specifically. the sound spectrogram has been a basic tool for gaining understanding of how the sounds of speech are produced and how phonetic information is encoded in the speech signal. In order to make smooth looking plots like those in Figure 4. R is usually quite small compared to both the window length L and the number of samples in the frequency dimension.6 are plots of S(tr. (4. and electrically sensitive paper [66]. where T is the sampling period of the discretetime signal x[n] = xa(nT).aliasing in frequency of the STFT. 4.5 The Speech Spectrogram Since the 1940s.
6 Spectrogram for speech signal of Figure 1.either a grayscale or a colormapped image. waveform at the top and two spectrograms computed with different .6 shows the time 4.1. Figure 4.5 The Speech Spectrogram 47 −1 0 1 SH UH D W IY CH EY S frequency in kHz 0 100 200 300 400 500 600 0 1 2 3 4 5 −100 −50 0 frequency in kHz time in ms 0 100 200 300 400 500 600 0 1 2 3 4 5 −120 −100 −80 −60 −40 −20 0 Fig. 4.
in voiced intervals. etc.6 was computed with a window length of 101 samples. The bars on the right calibrate the color map (in dB). which is characterized by good frequency resolution and poor time . The length of the window has a major effect on the spectrogram image. then mostly small amplitude 48 ShortTime Analysis of Speech samples. the spectrogram is a narrowband spectrogram. This is consistent with the linear filtering interpretation of the STFT. This window length is on the order of the length of a pitch period of the waveform during voiced intervals. corresponding to 10 ms time duration. the spectrogram is called a wideband spectrogram. when the window length is long.length analysis windows. Conversely. since a short analysis filter has a wide passband. For this reason.20) and the corresponding spectrogram images leads to valuable insight into the nature of the speech signal. First note that the window sequence w[m] is nonzero only over the interval 0 ≤ m ≤ L − 1. A careful interpretation of (4. if the analysis window is short. The upper spectrogram in Figure 4. each individual pitch period is resolved in the time dimension. the spectrogram displays vertically oriented striations corresponding to the fact that the sliding window sometimes includes mostly large amplitude samples. but the resolution in the frequency dimension is poor. As a result. As a result of the short analysis window.
corresponding to 40 ms time duration.5 5 −80 −60 −40 −20 0 frequency in kHz log magnitude (dB) Fig.5 2 2.5 4 4.6.5 5 −80 −60 −40 −20 0 log magnitude (dB) 0 0.7 Shorttime spectrum at time 430 ms (dark vertical line in Figure 4.6 Relation of STFT to STACF 49 be obtained at other times around tr = 430 ms. The lower spectrogram in Figure 4. Note the three broad peaks in the spectrum slice at time tr = 430 ms.5 1 1.5 4 4.7. 4.5 1 1. This vertical slice through the spectrogram is at the position of the black vertical line in the upper spectrogram of Figure 4.5 3 3. 4.5 2 2.5 3 3.resolution. shows S(tr. and observe that similar slices would 0 0. The upper plot in Figure 4. for example.6) with Hamming window of length M = 101 in upper plot and M = 401 in lower plot. These large peaks are representative of the underlying resonances of the vocal tract at the corresponding time in the production of the speech signal. This window length is on the order of several pitch periods of the waveform .6 was computed with a window length of 401 samples.fk) as a function of fk at time tr = 430 ms.
The lower plot in Figure 4. the spectrogram no longer displays vertically oriented striations since several periods are included in the window no matter where it is placed on the waveform in the vicinity of the analysis time tr. f0.7. but the resolution in the frequency dimension is much better. for example. and at integer multiples (harmonics) of the fundamental frequency [89].9 4. As a result.during voiced intervals. so over the analysis window interval it acts very much like a periodic signal.fk) as a function of fk at time tr = 430 ms. where T0 is the fundamental period (pitch period) of the signal. In this case. the spectrogram is not as sensitive to rapid time variations.6 Relation of STFT to STACF A basic property of Fourier transforms is that the inverse Fourier transform . the striations tend to be horizontally oriented in the narrowband case since the fundamental frequency and its harmonics are all resolved. where the local maxima of the curve are spaced at multiples of the fundamental frequency f0 = 1/T0. Periodic signals have Fourier spectra that are composed of impulses at the fundamental frequency. As a result.7. the signal is voiced at the position of the window. Multiplying by the analysis window in the timedomain results in convolution of the Fourier transform of the window with the impulses in the spectrum of the periodic signal [89]. This is evident in the lower plot of Figure 4. Therefore. shows S(tr.
an inverse DFT can be used to compute ˜φrR[_] = 1 N N_−1 k=0 ˜XrR(ej(2πk/N))2ej(2πk/N)_ (4. Figures 4..15) is a discretetime Fourier transform for fixed window position.N − L.8(c) show the voiced and unvoiced autocorrelation .8 illustrates the equivalence between the STACF and the STFT.L − 1 if N ≥ 2L.. (4.23) by setting _ = 0. but ˜φrR[_] = φrR[_] for _ = 0..14) shows that the shorttime energy can also be obtained from either (4. 50 ShortTime Analysis of Speech (4. If L<N <2L time aliasing will occur..1. Note that (4..22) The STACF can also be computed from the sampled STFT.1.8(a) and 4..9) is related to the STFT given by 9 Each “ripple” in the lower plot is essentially a frequencyshifted copy of the Fourier transform of the Hamming window used in the analysis.15) as φˆn[_] = 1 2π _π −π Xˆn(ej ˆω)2ej ˆω_dˆω. In particular..of the magnitudesquared of the Fourier transform of a signal is the autocorrelation function for that signal [89]. Figure 4.23) and ˜φrR[_] = φrR[_] for _ = 0. Since the STFT defined in (4.22) or (4.. it follows that φˆn[_] given by (4.
5(d).5 0 0.8(b) that there are approximately 18 local regularly spaced peaks in the range 0–2000 Hz. 4. On the right are the corresponding STFTs.5 1 0 10 20 30 40 −0. 4.5 1 0 2000 4000 6000 8000 −100 −50 0 50 0 2000 4000 6000 8000 −100 −50 0 50 (a) Voiced Autocorrelation (b) Voiced STFT (c) Unvoiced Autocorrelation (d) Unvoiced STFT time in msec frequency in kHz Fig. .5(b) and 4.functions that were shown in Figures 4.7 ShortTime Fourier Synthesis 51 the autocorrelation function of the voiced segment imply a fundamental frequency at this time of approximately 1000/9 = 111 Hz. The peaks around 9 and 18 ms in 0 10 20 30 40 −0.8 STACF and corresponding STFT. note in the STFT on the right in Figure 4. Thus.5 0 0. Similarly.
4. The left half of Figure 4.we can estimate the fundamental frequency to be 2000/18 = 111 Hz as before.9 shows a block diagram representation of the STFT as a combination of modulation.18) shows that XrR[k] is a downsampled output of the lowpass window filter with frequencydownshifted input x[n]e−j(2πk/N)n. The w[n] w[n] w[n] x x xx x x Short Time Modifications R R R R R R . followed by lowpass filtering. Alternatively.7 ShortTime Fourier Synthesis From the linear filtering point of view of the STFT.20) represent the downsampled (by the factor R) output of the process of bandpass filtering with impulse response w[n]ej(2πk/N)n followed by frequencydownshifting by (2πk/N). (4. (4. and the outputs of the individual filters are called the channel signals. This structure is often called a filter bank. followed by a downsampler by R.19) and (4.
4.9 Filterbank interpretation of shorttime Fourier analysis and synthesis. 52 ShortTime Analysis of Speech spectrogram is comprised of the set of outputs of the filter bank with each channel signal corresponding to a horizontal line in the spectrogram.f [n] f [n] f [n] + Xn[k] x[n] jNn eN ( 1) 2 − π − jNn eN ( 1) 2 π − jn eN (1) − 2π jn eN (0) − 2π jn eN (1) 2π jn eN (0) 2π y[n] XrR[k] YrR[k] Fig. we may need to consider if and how it is possible to reconstruct . If we want to use the STFT for other types of speech processing.
. interpolates the STFT YrR[k] to the time sampling rate of the speech signal for each channel k. The remainder of Figure 4.24). The remaining parts of the structure on the right in Figure 4. An example might be quantization of the channel signals for data compression. (4. This operation. which upshifts the lowpass interpolated channel signals back to their original frequency bands centered on frequencies (2πk/N). defined by the part within the parenthesis in (4. . This part of the diagram represents the synthesis equation y[n] = N_−1 k=0 _ ∞_ r=−∞ YrR[k]f[n − rR] ej(2πk/N)n. The modified STFT is denoted as YrR[k]. synthesis is achieved by modulation with ej(2πk/N)n. the diagram shows the possibility that the STFT might be modified by some sort of processing.10 Then. The sum of the upshifted signals is the synthesized output. First.9 shows how this can be done. from the channel signals.24) The steps represented by (4.the speech signal from the STFT.24) and the right half of Figure 4.9 implement the synthesis of a new time sequence y[n] from the STFT.e. i.9 involve first upsampling by R followed by linear filtering with f[n] (called the synthesis window).
It is often a scaled version of w[n]. and the analysis and synthesis can be accomplished using only real filtering and with modulation operations being implicitly achieved by the downsampling [24. it is most important to note that with careful choice of the parameters L.8 ShortTime Analysis is Fundamental to our Thinking 53 error (y[n] = x[n]) from the unmodified STFT YrR[k] = XrR[k]. special efficiencies result when L > R = N. However.25) Shorttime Fourier analysis and synthesis is generally formulated with equally spaced channels with equal bandwidths. Such filter banks can be implemented as tree structures where frequency bands are successively divided into low.There are many variations on the shorttime Fourier analysis/ synthesis paradigm. Such . The FFT algorithm can be used to implement both analysis and synthesis. One condition that guarantees exact reconstruction is [24. 4. it is possible to reconstruct the signal x[n] with negligible 10 The sequence f[n] is the impulse response of a LTI filter with gain R/N and normalized cutoff frequency π/R. R. However. 129]. ∞_ r=−∞ w[rR − n + qN]f[n − rR] = _ 1q=0 0 q _= 0. we have seen that models for auditory processing involve nonuniform filter banks. 129]. and N and careful design of the analysis window w[m] together with the synthesis window f[n].and HF bands. (4.
but first we will consider (in Chapter 5) the technique of cepstrum analysis.8 ShortTime Analysis is Fundamental to our Thinking It can be argued that the shorttime analysis principle. See. Much of our knowledge of perceptual effects is framed in terms of frequency analysis. models for auditory processing are based on a filter bank as the first stage of processing. We will have more to say about this in Chapter 7. which is based directly on the STFT. for example [18. The fact that almost perfect reconstruction can be achieved from the filter bank channel signals gives the shorttime Fourier representation major credibility in the digital speech processing tool kit. 4. 125.structures are essentially the same as wavelet decompositions. a topic beyond the scope of this text. as discussed briefly in Chapter 3. This importance is strengthened by the fact that. and the . the STFT representation provides a natural framework within which this knowledge can be represented and exploited to obtain efficient representations of speech and more general audio signals. 129]. and particularly the shorttime Fourier representation of speech. is fundamental to our thinking about the speech signal and it leads us to a wide variety of techniques for achieving our goal of moving from the sampled time waveform back along the speech chain toward the implicit message. and thus. but one for which a large body of literature exists.
Healy. more specifically. 5 Homomorphic Speech Analysis The STFT provides a useful framework for thinking about almost all the important techniques for analysis of speech that have been developed so far. which is based on the STACF and thus equivalently on the STFT. Oppenheim. and Stockham showed that the cepstrum . An important concept that flows directly from the STFT is the cepstrum.1 Definition of the Cepstrum and Complex Cepstrum The cepstrum was defined by Bogert. and that further Fourier analysis of the log spectrum can aid in detecting the presence 54 5. A more detailed discussion of the uses of the cepstrum in speech processing can be found in [110]. and Tukey to be the inverse Fourier transform of the log magnitude spectrum of a signal [16]. was motivated by the fact that the logarithm of the Fourier spectrum of a signal containing an echo has an additive periodic component depending only on the echo size and delay. 5. In this chapter. the shorttime cepstrum.technique of linear predictive analysis (in Chapter 6).1 Definition of the Cepstrum and Complex Cepstrum 55 of that echo. we explore the use of the shorttime cepstrum as a representation of speech and as a basis for estimating the parameters of the speech generation model. loosely framed in terms of spectrum analysis of analog signals. Their original definition. Schafer. of speech.
x[n]. (5. (5. Since we restrict our attention to real sequences. 109]. .1b) where log{X(ejω)} is the complex logarithm of X(ejω) defined as ˆX (ejω) = log{X(ejω)} = log X(ejω) + j arg[X(ejω)].is related to the more general concept of homomorphic filtering of signals that are combined by convolution [85. it follows from the symmetry properties of Fourier transforms that the cepstrum is the even part of the complex cepstrum.1c) The transformation implied by (5. They gave a definition of the cepstrum of a discretetime signal as c[n] = 1 2π _π −π log X(ejω)ejωndω. and they extended the concept by defining the complex cepstrum as ˆx[n] = 1 2π _π −π log{X(ejω)}ejωndω.1.1b) is depicted as the block diagram in Figure 5. c[n] = (ˆx[n] + ˆx[−n])/2 [89].. 90. i. (5.1a) where log X(ejω) is the logarithm of the magnitude of the DTFT of the signal. The same diagram represents the cepstrum if the complex logarithm is replaced by the logarithm of the magnitude of the DTFT.e.
Homomorphic filtering of convolved signals is achieved by forming a modified complex cepstrum ˆy[n] = g[n]ˆx[n]. then ˆx[n] = D∗{x1[n] ∗ x2[n]} = ˆx1[n] + ˆx2[n]. the inverse of the characteristic system exists as in Figure 5. This property.As shown in Figure 5. That is. the complex cepstrum operator transforms convolution into addition.) .3) where g[n] is a window (a “lifter” in the terminology of Bogert et al. 56 Homomorphic Speech Analysis Fig.2) Fig.2 The inverse of the characteristic system for convolution (inverse complex cepstrum). (5. (5. In the theory of homomorphic systems D∗{ } is called the characteristic system for convolution. which shows the reverse cascade of the inverses of the operators in Figure 5.1.1. is what makes the cepstrum and the complex cepstrum useful for speech analysis. 5. since our model for speech production involves convolution of the excitation with the vocal tract impulse response.1 Computing the complex cepstrum using the DTFT. In the case of the complex cepstrum. and our goal is often to separate the excitation signal from the vocal tract signal. 5.2. the operation of computing the complex cepstrum from the input can be denoted as ˆx[n] = D∗{x[n]}. The connection between the cepstrum concept and homomorphic filtering of convolved signals is that the complex cepstrum has the property that if x[n] = x1[n] ∗ x2[n]. which is true for both the cepstrum and the complex cepstrum.
i. the output of the inverse characteristic system will have the form y[n] = y1[n] ∗ y2[n]. etc. Bogert et al. which must be done so as to preserve an additive combination of phases for two signals combined by convolution [90. 109]..2 with ˆy[n] given by (5.e. To emphasize this interchanging of domains of reference.2.6. Therefore. They created many other special terms in this way including quefrency as the name .which selects a portion of the complex cepstrum for inverse processing.3) defines a linear operator in the conventional sense. more specifically.4 and 5. A modified output signal y[n] can then be obtained as the output of Figure 5. 5.3) as input. if ˆx[n] = ˆx1[n] + ˆx2[n] then ˆy[n] = g[n]ˆx1[n] + g[n]ˆx2[n]. Examples of lifters used in homomorphic filtering of speech are given in Sections 5. the computation of the phase angle arg[X(ejω)]. Observe that (5. The key issue in the definition and computation of the complex cepstrum is the computation of the complex logarithm. where ˆy1[n] = g[n]ˆx1[n] is the complex cepstrum of y1[n]. [16] coined the word cepstrum by transposing some of the letters in the word spectrum.2 The ShortTime Cepstrum 57 The independent variable of the cepstrum and complex cepstrum is nominally time. The crucial observation leading to the cepstrum concept is that the log spectrum can be treated as a waveform to be subjected to further Fourier analysis.
5. a “cepstrogram” would be an image obtained by plotting the magnitude of the shorttime cepstrum as a function of quefrency m and analysis time ˆn.for the independent variable of the cepstrum and liftering for the operation of linear filtering the log magnitude spectrum by the operation of (5. the cepstrum and complex cepstrum are defined in terms of the DTFT. but not 1 In cases where we wish to explicitly indicate analysistime dependence of the shorttime .3 Computation of the Cepstrum In (5.9) should be clear. The shorttime cepstrum is a sequence of cepstra of windowed finiteduration segments of the speech waveform. This is useful in the basic definitions. 5. (5.1b).15).3). and liftering are widely used today.1b). Thus the shorttime cepstrum is defined as cˆn[m] = 1 2π _π −π log Xˆn(ej ˆω)ej ˆωmdˆω. By analogy. Only the terms cepstrum.2 The ShortTime Cepstrum The application of these definitions to speech requires that the DTFT be replaced by the STFT. and the shorttime complex cepstrum is likewise defined by replacing X(ej ˆω) by Xˆn(ej ˆω) in (5.4) where Xˆn(ej ˆω) is the STFT defined in (4.1a) and (5.1 The similarity to the STACF defined by (4. quefrency.
but as in other instances.5c) Note the “tilde” ˜ symbol above ˆx[n] in (5.e.1 Computation Using the DFT Since the DFT (computed with an FFT algorithm) is a sampled (in frequency) version of the DTFT of a finitelength sequence (i.3.5b) ˜ˆ x[n] = 1 N N_−1 k=0 ˆX [k]ej(2πk/N)n.5a) ˆX [k] = log X[k] + j arg{X[k]} (5.3..4).cepstrum.5c) and in Figure 5. several computational options exist. Fortunately. (5. which shows that the complex cepstrum can be computed approximately using the equations X[k] = N_−1 n=0 x[n]e −j(2πk/N)n (5. we often suppress the subscript ˆn.1a) and (5. X[k] = X(ej2πk/N) [89]). 5. we will use ˆn for the analysis time and m for quefrency as in (5. 58 Homomorphic Speech Analysis for use in processing sampled speech signals. the DFT and inverse DFT can be substituted for the DTFT and its inverse in (5.1b) as shown in Figure 5.3. Its purpose is to emphasize that using the DFT instead of the DTFT .
. the phase of the sampled DTFT must be evaluated as samples of a continuous function of frequency. Furthermore. The effect of timedomain aliasing can be made negligible by using a large value for N. by the principal value. 127]. A variety of algorithms have been developed for phase unwrapping [90. 5. In order for the complex cepstrum to be evaluated properly. then it must be “unwrapped” modulo 2π in order to ensure that convolutions are transformed into additions.. A more serious problem in computation of the complex cepstrum is the computation of the complex logarithm. 109. it is not a problem in computing the cepstrum.1b). An identical equation holds for the timealiased cepstrum ˜c[n]. ˜ˆ x[n] = ∞_ r=−∞ ˆx[n + rN]. If the principal value phase is first computed at discrete frequencies.results in an approximation to (5.1b) due to the timedomain aliasing resulting from the sampling of the log of the DTFT [89].e. 5. (5. i. since the phase is not used. This is Fig. While accurate phase unwrapping presents a challenge in computing the complex cepstrum.6) where ˆx[n] is the complex cepstrum defined by (5.3 Computing the cepstrum or complex cepstrum using the DFT.3 Computation of the Cepstrum 59 because the angle of a complex number is usually specified modulo 2π. That is.
2 zTransform Analysis The characteristic system for convolution can also be represented by the twosided ztransform as depicted in Figure 5. 5. For this purpose. This is very useful for theoretical investigations.8a) Xuc(z) = M_uc k=1 (1 − ejθkz −1). (5. (5.3.4.7) where Xmax(z) = zMo _Mo k=1 (1 − akz −1) = _Mo k=1 (−ak) _Mo k=1 (1 − a −1 k z).phase unwrapping can be avoided by using numerical computation of the poles and zeros of the ztransform. (5.8b) Xmin(z) = A . we assume that the input signal x[n] has a rational ztransform of the form: X(z) = Xmax(z)Xuc(z)Xmin(z). and recent developments in polynomial root finding [117] have made the ztransform representation a viable computational basis as well.
i. respectively.4 ztransform representation of characteristic system for convolution...e. zk = ak. are zeros of X(z) outside of the unit circle (ak > 1). where bk and ck are zeros and poles. (5. i.11). 5. Xuc(z) contains all the zeros (with angles θk) on the unit circle. The minimumphase part is Xmin(z).8c) The zeros of Xmax(z). It is included to simplify the results in (5._Mi k=1 (1 − bkz −1) _Ni k=1 (1 − ckz −1) . The complex cepstrum of x[n] is determined by assuming that the complex logarithm log{X(z)} results in the sum of logarithms of each of the product terms. 60 Homomorphic Speech Analysis Fig. that are inside the unit circle (bk < 1 and ck < 1).e. ˆX (z) = log _____ _Mo k=1 (−ak) _____ . Xmax(z) is thus the maximumphase part of X(z). The factor zMo implies a shift of Mo samples to the left.
(5.+ _Mo k=1 log(1 − a −1 k z) + M_uc k=1 log(1 − ejθkz −1) + log A + _Mi k=1 log(1 − bkz −1) − _Ni k=1 log(1 − ckz −1).9) and collecting the coefficients of the positive and negative powers of z gives ˆx[n] = _Mo k=1 ank .9) Applying the power series expansion log(1 − a) = − ∞_ n=1 an n a < 1 (5.10) to each of the terms in (5.
This is 5. (5.3 Computation of the Cepstrum 61 the case in theoretical analysis where the poles and zeros are specified. .11) Given all the poles and zeros of a ztransform X(z).11) allows us to compute the complex cepstrum with no approximation. (5.n n<0 log A + log _____ _mo k=1 (−ak) _____ n=0 _ − M_uc k=1 ejθkn n − _Mi k=1 bnk n + _Ni k=1 cnk n n > 0.
12) A second method that yields a ztransform is the method of linear predictive analysis.However. signals having a ztransform whose poles and zeros are inside the unit circle.e. X(z) = M_ n=0 x[n]z −n = A _Mo k=1 (1 − akz −1) _Mi k=1 (1 − bkz −1). The ztransform is then simply a polynomial with the samples x[n] as coefficients. All that is needed is a process for obtaining the ztransform as a rational function and a process for finding the zeros of the numerator and denominator.. One method of obtaining a ztransform is simply to select a finitelength sequence of samples of a signal. i.3 Recursive Computation of the Complex Cepstrum Another approach to computing the complex cepstrum applies only to minimumphase signals. (5.11) is also useful as the basis for computation. 5.3. An example would be the impulse .. This has become more feasible with increasing computational power and with new advances in finding roots of large polynomials [117]. (5. i. to be discussed in Chapter 6.e.
In this case. From (5.14) 62 Homomorphic Speech Analysis It can be shown [90] that the impulse response and its complex cepstrum are related by the recursion formula: . (5.response of an allpole vocal tract model with system function H(z) = G 1− _p k=1 αkz −____________k =G _p k=1 (1 − ckz −1) .11) it follows that the complex cepstrum of the impulse response h[n] corresponding to H(z) is ˆh [n] = 0n<0 log G n = 0 _p k=1 cnk n n > 0. (5.13) Such models are implicit in the use of linear predictive analysis of speech (Chapter 6). all the poles ck must be inside the unit circle for stability of the system.
ˆh [n] = 0n<0 logG n= 0 αn + n_−1 k=1 _ k n .ˆh [n] = 0n<0 logG n= 0 h[n] h[0] − n_−1 k=1 _ k n _ˆh[k]h[n − k] h[0] n ≥ 1.e.13) and the complex cepstrum of the impulse response of the model filter.15) Furthermore. working with the reciprocal (negative logarithm) of (5. it can be shown that there is a direct recursive relationship between the coefficients of the denominator polynomial in (5. i.. (5.13).
16) it follows that the coefficients of the denominator polynomial can be obtained from the complex cepstrum through αn =ˆh[n] − n_−1 k=1 _ k n _ ˆh [k]αn−k 1 ≤ n ≤ p. This fact is the basis for the use of the cepstrum in speech coding and speech recognition as a vector representation of the vocal tract properties of a frame of speech. . .5 shows an example of the shorttime cepstrum of speech for the segments of voiced and unvoiced speech in Figures 4. 5.4 ShortTime Homomorphic Filtering of Speech Figure 5.5(a) and 4. (5.5(c).17). (5.17)._ ˆh [k]αn−k n > 0.13) since all the denominator coefficients and G can be computed from ˆh[n] for n= 0. it follows that p + 1 values of the complex cepstrum are sufficient to fully determine the speech model system in (5.16) From (5.17) From (5. . .1. p using (5. The low quefrency part of the cepstrum is expected to be representative .
high quefrency components would correspond to the more rapid fluctuations .of the slow variations (with frequency) in the log spectrum.4 ShortTime Homomorphic Filtering of Speech 63 0 10 20 30 40 −1 0 1 0 10 20 30 40 −1 0 1 0 2000 4000 6000 8000 −8 −6 −4 −2 0 2 4 0 2000 4000 6000 8000 −8 −6 −4 −2 0 2 4 (a) Voiced Cepstrum (b) Voiced log STFT (c) Unvoiced Cepstrum (d) Unvoiced log STFT time in ms frequency in kHz Fig.5 Shorttime cepstra and corresponding STFTs and homomorphicallysmoothed spectra. while the 5. 5.
Note that the spectrum for the voiced segment in Figure 5. The existence of this peak in the quefrency range of expected pitch periods strongly signals voiced speech. manifests itself in the cepstrum peak at a quefrency of about 9ms in Figure 5. This is illustrated by the plots in Figure 5.5(d). the autocorrelation function also displays an indication of periodicity. This periodic structure in the log spectrum of Figure 5. As shown in Figure 4.5(a) and 5. On the other hand. The log magnitudes of the corresponding shorttime spectra are shown on the right as Figures 5. Furthermore. As a result.of the log spectrum.5(c)). there is no strong peak indicating periodicity as in the voiced case.5(b).5.5(b) and 5. note that the rapid variations of the unvoiced spectra appear random with no periodic structure.5(b). . but not nearly as unambiguously as does the cepstrum. 64 Homomorphic Speech Analysis To illustrate the effect of liftering. the quefrency of the peak is an accurate estimate of the pitch period during the corresponding speech interval. This is typical of Fourier transforms (periodograms) of short segments of random signals.5(a). the quefrencies above 5ms are multiplied by zero and the quefrencies below 5 ms are multiplied by 1 (with a short transition taper as shown in Figures 5.5(b) has a structure of periodic ripples due to the harmonic structure of the quasiperiodic segment of voiced speech.
5(b) and 5. This is illustrated in Figure 5. it is apparent that for the positions 1 through 5. On the left is a sequence of log shorttime spectra (rapidly varying curves) and on the right is the corresponding sequence of cepstra computed from the log spectra on the left. From the discussion of Section 5.The DFT of the resulting modified cepstrum is plotted as the smooth curve that is superimposed on the shorttime spectra in Figures 5.5(d). the window includes only unvoiced .4. The successive spectra and cepstra are for 50 ms segments obtained by moving the window in steps of 12. Noll [83] applied the shorttime cepstrum to detect local periodicity (voiced speech) or the lack thereof (unvoiced speech). a useful perspective is that by liftering the cepstrum.6.5 ms (100 samples at a sampling rate of 8000 samples/sec). respectively. Therefore.5 Application to Pitch Detection The cepstrum was first applied in speech processing to determine the excitation parameters for the discretetime speech model of Figure 7. These slowly varying log spectra clearly retain the general spectral shape with peaks corresponding to the formant resonance structure for the segment of speech under analysis. it is possible to separate information about the vocal tract contributions from the shorttime speech spectrum [88]. 5.10. which shows a plot that is very similar to the plot first published by Noll [83].
while for positions 6 and 7 the signal within the window is partly voiced and partly unvoiced. As can be seen from the plots on the right. For positions 8 through 15 the window only includes voiced speech. Note again that the rapid variations of the unvoiced spectra appear random with no periodic structure.speech. and the quefrency of the peak is an accurate estimate of the pitch period during the corresponding speech interval. the spectra for voiced segments have a structure of periodic ripples due to the harmonic structure of the quasiperiodic segment of voiced speech. the cepstrum peak at a quefrency of about 11–12 ms strongly signals voiced speech. 5.5 Application to Pitch Detection 65 01234 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0 5 10 15 20 25 1 2 . On the other hand.
5. the pitch detection algorithm includes many features designed to handle cases that do not fit the underlying model very well. For example.3 4 5 6 7 8 9 10 11 12 13 14 15 ShortTime Log Spectra in Cepstrum Analysis frequency in kHz ShortTime Cepstra time in ms window number Fig. and the quefrency location of the peak gives the estimate of the pitch period. Presence of a strong peak implies voiced speech.6 Shorttime cepstra and corresponding STFTs and homomorphicallysmoothed spectra. The essence of the pitch detection algorithm proposed by Noll is to compute a sequence of shorttime cepstra and search each successive cepstrum for a peak in the quefrency region of the expected pitch period. As in most modelbased signal processing applications of concepts such as the cepstrum. for frames 6 and 7 the cepstrum peak is weak 66 Homomorphic Speech Analysis .
Such comparisons require a distance measure. In either case. . cepstra can be computed either by ztransform analysis or by DFT implementation of the characteristic system. a test pattern c[n] (vector of cepstrum values n = 1.corresponding to the transition from unvoiced to voiced speech. In other problematic cases. . As we have shown. In such applications. nco) is compared against a similarly defined reference pattern ￣c[n]. we can assume that the cepstrum vector corresponds to a gainnormalized (c[0] = 0) minimumphase vocal tract impulse response that is defined by the complex cepstrum2 ˆh [n] = _ 2c[n] 1≤ n ≤ nco 0 n < 0. a speech signal is represented on a framebyframe basis by a sequence of shorttime cepstra.18) In problems such as VQ or ASR.6 Applications to Pattern Recognition Perhaps the most pervasive application of the cepstrum in speech processing is its use in pattern recognition systems such as design of vector quantizers (VQ) and automatic speech recognizers (ASR). 5. . the peak at twice the pitch period may be stronger than the peak at the quefrency of the pitch period. . . (5.2. Noll applied temporal continuity constraints to prevent such errors.
(5. Therefore. Thus. the Euclidean distance applied to the cepstrum would give D= _nco n=1 c[n] − ￣c[n]2 . 5.For example. cepstrumbased comparisons are strongly related to comparisons of smoothed shorttime spectra. and its interpretation as a difference of log spectra suggests a match to auditory perception mechanisms.18).18) or the real part of the 2For minimumphase signals. it follows that ˆh[n] = 2c[n] for n > 0. .19a) Equivalently in the frequency domain. (5. Since the cepstrum is always the even part of the complex cepstrum. the cepstrum offers an effective and flexible representation of speech for pattern recognition problems. D= 1 2π _π −π ___ log H(ej ˆω) − log ￣H (ej ˆω) ___ 2 dˆω.6 Applications to Pattern Recognition 67 DTFT ofˆh[n] in (5. the complex cepstrum satisfies h[n] = 0 for n < 0.19b) where log H(ej ˆω) is the log magnitude of the DTFT of h[n] corresponding to the complex cepstrum in (5.
i. the shorttime cepstrum of one frame of the filtered speech signal y[n] will be3 c (y) ˆn [m] = c (x) ˆn [m] + c(h)[m]. c (x) n [m] = c (y) ˆn [m] − c(h)[m]. y[n] = h[n] ∗ x[n].1 Compensation for Linear Filtering Suppose that we have only a linearly filtered version of the speech signal. instead of x[n]. if we can estimate c(h)[n]. Another approach to removing the effects of linear distortions is to .4 which we assume is nontimevarying. This property is extremely attractive in situations where the reference pattern ￣c[m] has been obtained under different recording or transmission conditions from those used to acquire the test vector.6.5. the test vector can be compensated for the effects of the linear filtering prior to computing the distance measures used for comparison of patterns.e. If the analysis window is long compared to the length of h[n]. we can obtain c (x) ˆn [m] at each frame from c (y) ˆn [m] by subtraction. (5. In these circumstances..20) where c(h)[m] will appear essentially the same in each frame. Therefore.
where m denotes the quefrency index of the cepstrum. Therefore.observe that the cepstrum component due to the distortion is the same in each frame. the linear distortion effects are removed. 4 Stockham [124] showed how c(h)[m] for such linear distortions can be estimated from the signal y[n] by time averaging the log of the STFT. 3 In this section. we denote the cepstrum at analysis time ˆn of a signal x[n] as c (x) ˆn [m]. (5. i. Specifically. 40] first noted that the sequence of cepstrum values has . it will be useful to use somewhat more complicated notation. it can be removed by a simple first difference operation of the form: Δc (y) ˆn [m] = c (y) ˆn [m] − c (y) ˆn−1[m].e.21) It is clear that if c (y) ˆn [m] = c (x) ˆn [m] + c(h)[m] with c(h)[m] being independent of ˆn. 68 Homomorphic Speech Analysis Furui [39. then Δc (y) ˆn [m] = Δc (x) ˆn [m]..
and additive noise [63. that when averaged over many frames of speech and speakers. (5. .21) is simply the slope of a firstorder polynomial fit to the cepstrum time evolution. cepstrum values c[n] have zero means and variances on the order of 1/n2. bias toward harmonic peaks. it is observed that there is significant statistical variability due to a variety of factors including shorttime analysis window position. 126]. He used polynomial fits to cepstrum sequences to extract simple representations of the temporal variation.22a) which can be written as the Euclidean distance of liftered cepstra D= _nco n=1 g[n]c[n] − g[n]￣c[n]2 . .6. The delta cepstrum as defined in (5.temporal information that could be of value for a speaker verification system. (5. .2.2 Liftered Cepstrum Distance Measures In using linear predictive analysis (discussed in Chapter 6) to obtain cepstrum feature vectors for pattern recognition problems. nco could be used to equalize the contributions for each term to the cepstrum difference. This suggests that g[n] = n for n = 1.22b) Tohkura [126] found. . A solution to this problem is to use weighted distance measures of the form: D= _nco n=1 g2[n] c[n] − ￣c[n]2 . 5. for example. .
This was attributed to the fact that for larger s the group delay spectrum becomes very sharply peaked and thus more sensitive to small differences in formant locations. Instead of g[n] = n for all n.6 Applications to Pattern Recognition 69 the lifter of (5. (5.23). (5. For example.Juang et al.5nco sin(πn/nco) n = 1.23) Tests of weighted distance measures showed consistent improvements in automatic speech recognition tasks. If s = 1 and τ is large. . . or 5. if s = 0 we have simply lowpass liftering of the cepstrum. . recognition rates improved significantly with τ = 5 and increasing values of the parameter s. They found that for clean test utterances. However. nco. we have essentially g[n] = n for small n with high quefrency tapering.24) This lifter has great flexibility. the difference in recognition rate was small for different values of s when τ ≈ 5 although performance suffered with increasing s for larger values of τ . . Itakura and Umezaki [55] tested the group delay spectrum distance measure in an automatic speech recognition system. .2. Itakura proposed the lifter g[n] = nse −n2/2τ2 . Itakura and Umezaki [55] used the group delay function to derive a different cepstrum weighting function. in test conditions with additive white noise and also with linear filtering distortions. [63] observed that the variability due to the vagaries of LPC analysis could be lessened by using a lifter of the form: g[n]=1 + 0.
The basic idea is to compute a frequency analysis based upon a filter bank with approximately critical band spacing of the filters and bandwidths. as noted in Chapter 3. With this in mind. For 4 kHz bandwidth. Then the DFT values are grouped together in critical bands and weighted by a triangular weighting function as depicted in Figure 5. which. Note that the bandwidths in Figure 5.5. This is significant in light of models for human perception of sound. In most implementations. approximately 20 filters are used. are based upon a frequency analysis performed in the inner ear.7 are constant for center frequencies below 1 kHz and then increase exponentially up to half the sampling rate of 4 kHz resulting 70 Homomorphic Speech Analysis 0 1000 2000 3000 4000 0 0.005 .7.6. a shorttime Fourier analysis is done first.3 MelFrequency Cepstrum Coefficients As we have seen. weighted cepstrum distance measures have a directly equivalent interpretation in terms of distance in the frequency domain. resulting in a DFT Xˆn[k] for analysis time ˆn. Davis and Mermelstein [27] formulated a new type of cepstrum representation that has come to be widely used and known as the melfrequency cepstrum coefficients (mfcc).
0.. in a total of 22 “filters. a discrete cosine transform of the log of the magnitude of the filter outputs is computed to form the function mfccˆn[m].25b) is a normalizing factor for the rth melfilter.” The melfrequency spectrum at analysis time ˆn is defined for r = 1.7. where Ar = _Ur k=Lr Vr[k]2 (5. 5.. It is needed so that a perfectly flat input Fourier spectrum will produce a flat melspectrum. For each frame.25a) where Vr[k] is the triangular weighting function for the rth filter ranging from DFT index Lr to Ur.2.7 Weighting functions for Melfrequency filter bank. i.. (5.e.R as MFˆn[r] = 1 Ar _Ur k=Lr Vr[k]Xˆn[k]2 . This normalization is built into the weighting functions of Figure 5. mfccˆn[m] = 1 R _R r=1 log(MFˆn[r])cos _ 2π ..01 frequency in Hz Fig..
The large dots are the values of log(MFˆn[r]) and the line interpolated 5.8 shows the result of mfcc analysis of a frame of voiced speech in comparison with the shorttime Fourier spectrum. Nmfcc = 13 and R = 22.R _ r+ 1 2 _ m _ . Figure 5. (5. LPC spectrum (discussed in Chapter 6). that is less than the number of melfilters. e.26) Typically. Nmfcc.g. and a homomorphically smoothed spectrum.7 The Role of the Cepstrum 71 0 500 1000 1500 2000 2500 3000 3500 4000 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 log magnitude frequency in Hz .. mfccˆn[m] is evaluated for a number of coefficients.
the cepstrum entered the realm of speech processing as a basis for pitch detection. Note that the delta cepstrum idea expressed by (5. the reconstructed melspectrum of course has more smoothing due to the structure of the filter bank. 111]. 5. the linear predictive analysis methods to be discussed in the next chapter have proven to be . Mel cepstrum smoothing. 88. 5.7 The Role of the Cepstrum As discussed in this chapter. between them is a spectrum reconstructed at the original DFT frequencies.Shorttime Fourier Transform Homomorphic smoothing. It remains one of the most effective indicators of voice pitch that have been devised.8 Comparison of spectral smoothing methods. While separation techniques based on the cepstrum can be very effective. Because the 72 Homomorphic Speech Analysis vocal tract and excitation components are well separated in the cepstrum. Fig. At higher frequencies. it was natural to consider analysis techniques for estimation of the vocal tract system as well [86. Note that all these spectra are different. LPC smoothing.21) can be applied to mfcc to remove the effects of linear filtering as long as the frequency response of the distorting linear filter does not vary much across each of the melfrequency bands. but they have in common that they have peaks at the formant resonances.
1. or random noise (during unvoiced speech). Nevertheless. 6. 6 Linear Predictive Analysis Linear predictive analysis is one of the most powerful and widely used speech analysis techniques. where the speech model is the part inside the dashed box. Over short time intervals. and discuss some of the issues that are involved in using it in practical speech applications.1 Model for linear predictive analysis of speech signals. where the sampled speech signal was modeled as the output of a linear. the linear system is 73 74 Linear Predictive Analysis Fig.1 Linear Prediction and the Speech Model We come to the idea of linear prediction of speech by recalling the source/system model that was introduced in Chapter 2. described by an allpole system function of the form: H(z) = S(z) . The particular form of the source/system model implied by linear predictive analysis is depicted in Figure 6. the cepstrum concept has demonstrated its value when applied to vocal tract system estimates obtained by linear predictive analysis. In this chapter. slowly timevarying system excited by either quasiperiodic impulses (during voiced speech). The importance of this method lies both in its ability to provide accurate estimates of the speech parameters and in its relative speed of computation.more effective for a variety of reasons. we present a formulation of the ideas behind linear prediction. 6.
and the filter coefficients {ak} can be estimated in a very straightforward and computationally efficient manner by the method of linear predictive analysis. i. For the system of Figure 6. the excitation is whatever is needed to produce s[n] at the output of the system. is defined as a system whose output is ˜s[n] = _p k=1 αks[n − k].1 with the vocal tract model of (6.3) and the prediction error.E(z) =G 1− _p k=1 akz −k . (6. the excitation is defined implicitly by the vocal tract system model. defined as the amount by which ˜s[n] fails to exactly predict sample s[n]. αk.1) In linear predictive analysis.1).e. (6. The major advantage of this model is that the gain parameter. is d[n] = s[n] − ˜s[n] = s[n] − ..2) A linear predictor with prediction coefficients. (6. G. the speech samples s[n] are related to the excitation e[n] by the difference equation s[n] = _p k=1 aks[n − k] + Ge[n].
4) that if the speech signal obeys the model of (6. The resulting parameters are then assumed to be the parameters of the system function H(z) in the model for production of the given segment of the speech waveform. then d[n] = Ge[n]._p k=1 αks[n − k]. i. of (6.4) it follows that the prediction error sequence is the output of an FIR linear system whose system function is A(z) = 1 − _p k=1 αkz −k = D(z) S(z) .1).e. Thus. (6. H(z) = G A(z) .5) It can be seen by comparing (6. The basic approach is to find a set of predictor coefficients that will minimize the meansquared prediction error over a short segment of the speech waveform.2) exactly.1 Linear Prediction and the Speech Model 75 From (6.2) and (6.. H(z). This process is repeated periodically at a rate appropriate .4) 6.6) The basic problem of linear prediction analysis is to determine the set of predictor coefficients {αk} directly from the speech signal in order to obtain a useful estimate of the timevarying vocal tract system through the use of (6. will be an inverse filter for the system. and if αk = ak. (6. A(z). (6. the prediction error filter.6).
P0.. then it can be shown that the predictor coefficients that result from minimizing the meansquared prediction error (over all time) are identical to the coefficients of (6. A third pragmatic justification for using the minimum meansquared prediction error as a 76 Linear Predictive Analysis basis for estimating the model parameters is that this approach leads to an exceedingly useful and accurate representation of the speech signal that can be obtained by efficient solution of a set of linear equations.2) with nontimevarying coefficients and excited either by a single impulse or by a stationary white noise input.to track the phonetic variation of speech (i. For voiced speech this means that d[n] would consist of a train of impulses. First. order of 50–100 times per second). then d[n] = Ge[n].. Thus. but it can be justified in several ways.e. That this approach will lead to useful results may not be immediately obvious. A second motivation for this approach follows from the fact that if a signal is generated by (6. recall that if αk = ak. finding αks that minimize the meansquared prediction error seems consistent with this observation. d[n] would be small except at isolated samples spaced by the current pitch period.e. i.2). The shorttime average prediction error is defined as .
2. sˆn[m] = s[m + ˆn] − M1 ≤ m ≤ M2.7).9) can be written more compactly as2 .7) by setting ∂Eˆn/∂αi = 0. (6. (6.. The details of specific definitions of the averaging operation will be discussed below. We can find the values of αk that minimize Eˆn in (6. .7) where sˆn[m] is a segment of speech that has been selected in a neighborhood of the analysis time ˆn. .10) then (6. (6. The notation _ _ denotes averaging over a finite number of samples. (6.9) where the ˜αk are the values of αk that minimize Eˆn in (6. the time origin of the analysis segment is shifted to sample ˆn of the entire signal. (Since the ˜αk are unique.k] = _sˆn[m − i]sˆn[m − k]_ . .e.) If we define ϕˆn[i.8) That is. . we will drop the tilde and use the notation αk to denote the values that minimize Eˆn. i. thereby obtaining the equations1 _p k=1 ˜αk _sˆn[m − i]sˆn[m − k]_ = _sˆn[m − i]sˆn[m]_ 1 ≤ i ≤ p. for i = 1. p.Eˆn = _ d2 ˆn[m] _ = __ sˆn[m] − _p k=1 αksˆn[m − k] 2_ .
. . . the total minimum meansquared error consists of a fixed component equal to the meansquared value of the signal segment minus a term that depends on the predictor coefficients that satisfy (6. (6. the minimum meansquared prediction error can be shown to be [3.2 Computing the Prediction Coefficients 77 If we know ϕˆn[i.11). i.13) the most.e. (6. 0] i = 1.k] for 1 ≤ i ≤ p and 0 ≤ k ≤ p. . The details of the definition of the averaging operation used in (6.10) have a significant effect on the properties of the prediction coefficients that are obtained by solving (6. the solutions of (6.2. 2 The quantities ϕˆn[i. .9).7) only provide a stationary point which can be shown to be a minimum of Eˆn since Eˆn is a convex function of the αis.12) can be solved for the vector α = {αk} of unknown predictor coefficients that minimize the average squared prediction error for the segment sˆn[m]. (6. this set of p equations in p unknowns.k] = ϕˆn[i.11) 1 More accurately.13) Thus.11)..k] are in the form of a correlation function for the speech segment sˆn[m]._p k=1 αkϕˆn[i.k].3 Using (6. p. which can be represented by the matrix equation: Φα = ψ. 0] − _p k=1 αkϕˆn[0.7) and (6. the optimum coefficients reduce Eˆn in (6. 5] Eˆn = ϕˆn[0. 6.
the details of the computation of ϕˆn[i. We shall see below that two methods for linear predictive analysis emerge out of a consideration of the limits of summation and the definition of the waveform segment sˆn[m].2. and ϕˆn[i.k] for 1 ≤ i ≤ p and 0 ≤ k ≤ p. Once this is done we only have to solve (6. 6.11) to obtain the αks. in a shorttime analysis procedure.To solve for the optimum predictor coefficients.k] and the subsequent solution of the equations are somewhat intricate and further discussion is required. applied to the same speech signal yield slightly different optimum predictor coefficients. However. Thus. the averaging must be over a finite interval.4 3 Although the αks are functions of ˆn (the time index at which they are estimated) this dependence will not be explicitly shown. linear prediction analysis is very straightforward. 4 These two methods.7).2 Computing the Prediction Coefficients So far we have not been explicit about the meaning of the averaging notation _ _ used to define the meansquared prediction error in (6. sˆn[m]. As we have stated.k] when no confusion will result.1 The Covariance Method One approach to computing the prediction coefficients is based on the . we must first compute the quantities ϕˆn[i. We shall also find it advantageous to drop the subscripts ˆn on Eˆn. 78 Linear Predictive Analysis 6. in principle.
11) inherit the same definition of the averaging operator. (6. (6. (6.15) require values of sˆn[m] = s[m + ˆn] over the range −M1 − p ≤ m ≤ M2. ϕˆn[i.k] = M_2−i m=−M1−i sˆn[m]sˆn[m + i − k] (6. The quantities ϕˆn[i. By changes of index of summation..14) and (6.16a) = .k] = _M2 m=−M1 sˆn[m − i]sˆn[m − k] _ 1≤i≤p 0 ≤ k ≤ p. i.14) where −M1 ≤ n ≤ M2.k] needed in (6.e.15) can be expressed in the equivalent forms: ϕˆn[i.15) Both (6.definition Eˆn = _M2 m=−M1 (dˆn[m])2 = _M2 m=−M1 _ sˆn[m] − _p k=1 αksˆn[m − k] 2 .
The third plot shows the prediction error computed with the optimum predictor coefficients.11).2 Illustration of windowing and prediction error for the covariance method.M_2−k m=−M1−k sˆn[m]sˆn[m + k − i].2 Computing the Prediction Coefficients 79 Fig.14) and (6.k] for all required values of i and k.k] = ϕˆn[k. Figure 6.15).14). Note that in solving for the optimum prediction coefficients using (6. (6. The second plot shows that segment extracted as a finitelength sequence sˆn[m] = s[m + ˆn] for −M1 − p ≤ m ≤ M2. i]. . The minimum meansquared prediction error would be simply the sum of squares of all the samples shown. the 6. and the box denotes a segment of that waveform selected around some time index ˆn.16b) from which it follows that ϕˆn[i. This method does not require any assumption about the signal outside the interval −M1 − p ≤ m ≤ M2 since the samples sˆn[m] for −M1 − p ≤ m ≤ M2 are sufficient to evaluate ϕ[i. Note the p “extra” samples (light shading) at the beginning that are needed to start the prediction error filter at time −M1. 6.2 shows the sequences that are involved in computing the meansquared prediction error as defined by (6. The top part of this figure shows a sampled speech signal s[m]. prediction error is implicitly computed over the range −M1 ≤ m ≤ M2 as required in (6.
1) The meansquared prediction error satisfies Eˆn ≥ 0. (C. the prediction error sequence is not explicitly computed since solution of (6. (6.12) is a symmetric positivesemidefinite matrix.2) The matrix Φ in (6.11) reduces to the STACF φˆn[i − k] that we discussed in Chapter 4 [53.6) is not guaranteed to be stable. the well known Cholesky decomposition of the covariance matrix Φ [3]. This implies that the vocal tract model filter (6. the equations (6. the analysis segment sˆn[m] is defined as sˆn[m] = _ s[n + m]w[m] −M1 ≤ m ≤ M2 0 otherwise. In the autocorrelation method.3) The roots of the prediction error filter A(z) in (6. 80 Linear Predictive Analysis (C.2 The Autocorrelation Method Perhaps the most widely used method of linear predictive analysis is called the autocorrelation method because the covariance function ϕˆn[i. 6. 74. (C.2. it is theoretically possible for the average error to be exactly zero.17) where the analysis window w[m] is used to taper the edges of the segment . 78]. It is worthwhile to summarize them as follows: (C.k] needed in (6. The mathematical structure that defines the covariance method of linear predictive analysis implies a number of useful properties of the solution.2). for example. 54. With this method.5) are not guaranteed to lie within the unit circle of the zplane.4) As a result of (C.11) can be solved efficiently using.In most cases of linear prediction.11) does not require it.
k] is a function only of i − k. it follows that the prediction error sequence dˆn[m] can be nonzero only in the range −M1 ≤ m ≤ M2 + p. p. Applying this notion to (6. we can replace ϕˆn[i. Therefore.16a) and (6. . (6. Eˆn is defined as Eˆn = M_2+p m=−M1 (dˆn[m])2 = ∞_ m=−∞ (dˆn[m])2.2.to zero.20) The resulting set of equations for the optimum predictor coefficients is therefore _p k=1 αkφˆn[i − k] = φˆn[i] i = 1.18) The windowing of (6.17) to be zero outside the interval −M1 ≤ m ≤ M2. (6. ϕ[i.21) .16b) leads to the conclusion that ϕˆn[i. Therefore. (6. which is the STACF defined in Chapter 4 as φˆn[k] = ∞_ m=−∞ sˆn[m]sˆn[m + k] = φˆn[−k]. (6.k] by φˆn[i − k]. .k] = ∞_ m=−∞ sˆn[m]sˆn[m + i − k] = φˆn[i − k]. .17) allows us to use the infinite limits to signify that the sum is over all nonzero values of dˆn[m].19) Thus. . Since the analysis segment is defined by the windowing of (6.
Figure 6. Note that the zerovalued samples outside the window are shown with light shading. The upper plot shows the same sampled speech signal s[m] as in Figure 6. It is easy to see that at least one of these first p samples of the prediction error must be nonzero. The middle plot shows the result of multiplying the signal s[ˆn + m] by the window w[m] and redefining the time origin to obtain sˆn[m].2 Computing the Prediction Coefficients 81 Fig.2 with a Hamming window centered at time index ˆn. The third plot shows the prediction error computed using the optimum coefficients. 6. Also note the lightly shaded p samples at the beginning. Note that for this segment.3 shows the sequences that are involved in computing the optimum prediction coefficients using the autocorrelation method. Similarly.3 Illustration of windowing and prediction error for the autocorrelation method. It can also be seen that at least one of these last p samples of the prediction error . the prediction error (which is implicit in the solution of (6.6.21)) is nonzero over the range −M1 ≤ m≤ M2 + p. the last p samples of the prediction error can be large due to the fact that the predictor must predict zerovalued samples from windowed speech samples. These samples can be large due to the fact that the predictor must predict these samples from the zerovalued samples that precede the windowed segment sˆn[m].
the equations (6. because of its many implications. must always be strictly greater than zero. With this method.2). it follows that Eˆn. 6. Equation (6.1) The meansquared prediction error satisfies Eˆn > 0. which.6) is guaranteed to be stable. being the sum of 82 Linear Predictive Analysis squares of the prediction error samples.2) The matrix Φ in (6.11) can be solved efficiently using the Levinson–Durbin algorithm.3) The roots of the prediction error filter A(z) in (6.5) are guaranteed to lie within the unit circle of the zplane so that the vocal tract model filter of (6.3 The Levinson–Durbin Recursion As stated in (A.3. For this reason.12) is a symmetric positivedefinite Toeplitz matrix.22) shows the detailed structure of the matrix equation Φα = ψ for the autocorrelation method.must be nonzero. the mathematical structure of the autocorrelation method implies a number of properties of the solution.5 As in the case of the covariance method. (A.2) above. it is theoretically impossible for the error to be exactly zero because there will always be at least one sample at the beginning and one at the end of the prediction error sequence that will be nonzero. . which means that all the elements on a given diagonal in the matrix are equal.12) is a symmetric positivedefinite Toeplitz matrix [46]. including the following: (A. (A. the matrix Φ in (6.4) As a result of (A. (A. we discuss in more detail in Section 6.
3 The Levinson–Durbin Recursion 83 Note that the vector ψ is composed of almost the same autocorrelation values as comprise Φ. Because of the special structure of (6.22) 5For this reason. a tapering window is generally used in the autocorrelation method. That algorithm.φ[0] φ[1] · · · φ[p − 1] φ[1] φ[0] · · · φ[p − 2] ············ φ[p − 1] φ[p − 2] · · · φ[0] α1 α2 ··· αp = φ[1] φ[2] ··· φ[p] (6. known as the Levinson–Durbin algorithm.22) it is possible to derive a recursive algorithm for inverting the matrix Φ. is specified by the following steps: . 6.
2) α (i) i = ki (D. i − 1 α (i) j=α (i−1) j − kiα (i−1) i−j (D.1) for i = 1. . . . p (D.3) if i > 1 then for j = 1. . . . p ki = φ[i] − _i−1 j=1 α (i−1) j φ[i − j] E(i−1) (D.2.5) end αj = α (p) j j = 1. .2.2. . . . .6) An important feature of the Levinson–Durbin algorithm is that it . .4) end E(i) = (1 − k2 i )E(i−1) (D.Levinson–Durbin Algorithm E0 = φ[0] (D.
(6. all predictors from order 0 (no prediction) to order p are computed along with the corresponding meansquared prediction errors E(i).3) and (D. (b) Vocal tract filter H(z) = 1/A(z). 84 Linear Predictive Analysis Fig. it is possible to start at the output of Figure 6. (a) Prediction error filter A(z). 6. and as part of the process.23) leads (after some manipulations) to an interpretation of the Levinson–Durbin algorithm in terms of a lattice filter structure as in Figure 6. Also. Specifically.4(b). the equations labeled (D.4(a) and work to the left eventually computing s[n] in terms of e[n].4) can be used to show that the prediction error system function satisfies A(i)(z) = A(i−1)(z) − kiz −iA(i−1)(z −1). If the input and output signals in such physical models are sampled at just the right sampling .determines by recursion the optimum ithorder predictor from the optimum (i − 1)thorder predictor. Lattice structures like this can be derived from acoustic principles applied to a physical model composed of concatenated lossless tubes [101].4 Lattice structures derived from the Levinson–Durbin recursion. by solving two equations in two unknowns recursively. (6.4(a).23) Defining the ithorder forward prediction error e(i)[n] as the output of the prediction error filter with system function A(i)(z) and b(i)[n] as the output of the ithorder backward prediction error filter B(i)(z) = z−iA(i)(z−1). The lattice structure corresponding to this is shown in Figure 6.
78. showed that the parameters ki in the Levinson–Durbin recursion and the lattice filter interpretation obtained from it also could be derived by looking at linear predictive analysis from a statistical perspective.. PARCOR (for partial correlation) coefficients [54]. They called the ki parameters. . the coefficients ki behave as reflection coefficients at the tube boundaries [3.rate. p play a key role in the Levinson–Durbin recursion and also in the lattice filter interpretation. i.4 LPC Spectrum 85 backward prediction errors at the output of the (i − 1)th stage of prediction in Figure 6. In this case. .e. 54]. Note that the parameters ki for i = 1. Itakura and Saito [53. the sampled signals are related by a transfer function identical to (6.2.6). ki = ∞_ m=−∞ e(i−1)[m]b(i−1)[m − 1] _ ∞_ m=−∞ ! e(i−1)[m] "2 ∞_ m=−∞ ! .4(a). because they can be computed directly as a ratio of crosscorrelation values between the forward and 6. . 101]. .
it follows that.4 LPC Spectrum The frequencydomain interpretation of linear predictive analysis provides an informative link to our earlier discussions of the STFT and cepstrum .24) In the PARCOR interpretation. it must be true that −1 < ki < 1 for all i.24) are identical to the ki obtained as a result of the Levinson–Durbin algorithm.24). since E(i) = (1 − k2 i )E(i−1) is strictly greater than zero for predictors of all orders. Specifically.5) of the algorithm.2) in the Levinson– Durbin algorithm can be replaced by (6. It can be shown that this condition on the PARCORs also guarantees that all the zeros of a prediction error filter A(i)(z) of any order must be strictly inside the unit circle of the zplane [74. from Equation (D. 6.b(i−1)[m − 1] "2 1/2 . Indeed. The PARCOR coefficients computed using (6.4(a) removes part of the correlation in the input signal. Equation (D. The Levinson–Durbin formulation provides one more piece of useful information about the PARCOR coefficients. (6. 78]. each stage of Figure 6. and the result is an algorithm for transforming the PARCOR representation into the linear predictor coefficient representation.
5(a) and 6. p are used to compute the prediction coefficients and gain. . obtained by evaluating H(z) on the unit circle at angles 2πf/fs. Figures 6. (6.6). with the first 23 values plotted with a heavy line.analysis. Sˆn(ej ˆω)2. φˆn[m]. which is the inverse discrete Fourier transform of the magnitudesquared of the STFT. the magnitudesquared of the frequency response of this system.5(c) show the STACF.5 shows a comparison between shorttime Fourier analysis and linear predictive spectrum analysis for segments of voiced and unvoiced speech. which in turn define the vocal tract system function H(z) in (6.1. of the windowed speech signal sˆn[m] = s[n + m]w[m]. . The values φˆn[m] for m = 0.25) and can be thought of as an alternative shorttime spectral representation. . Therefore. The autocorrelation method is based on the shorttime autocorrelation function. 86 Linear Predictive Analysis is of the form: H(ej2πf/fs)2 = __________ G 1− _p k=1 αke −j2πf/fs __________ 2 . . Figure 6. These values are used to compute the predictor coefficients and gain for an LPC model with .
while the overall shape is assumed to be determined by the effects of glottal pulse.5 1 0 5 10 15 −0. and radiation. As we have observed before.p = 22.5 0 0. The frequency responses of the corresponding vocal tract system models are computed using (6.25) where the sampling frequency is fs = 16 kHz. In 0 5 10 15 −0. vocal tract transfer function.5 0 0.5 1 0 2000 4000 6000 8000 −100 −50 0 50 0 2000 4000 6000 8000 −100 −50 0 50 (a) Voiced Autocorrelation (b) Voiced LPC Spectrum (c) Unvoiced Autocorrelation (d) Unvoiced LPC Spectrum time in ms frequency in kHz . These frequency responses are superimposed on the corresponding STFTs (shown in gray). the rapid variations with frequency in the STFT are due primarily to the excitation.
the excitation effects are removed by focusing on the lowtime autocorrelation coefficients.4 LPC Spectrum 87 cepstrum analysis. the excitation effects are removed by lowpass liftering the cepstrum. The question naturally arises as to how p should be chosen. 88 Linear Predictive Analysis smoothing of the STFT. and this is exactly what is desired. That this is true in general can be argued using the Parseval theorem of Fourier analysis [74]. In Figure 6. the linear predictive spectra tend to favor the peaks of the shorttime Fourier transform. The amount of smoothing of the spectrum is controlled by the choice of p. and (b) the normalized meansquared prediction error for the same signal. but does not represent all its local peaks and valleys.5 shows that a linear prediction model with p = 22 matches the general shape of the shorttime spectrum. In linear predictive spectrum analysis. 6. which tends toward an average of the peaks . Figure 6. 6.Fig. Figure 6. 6. Evidently.6(a) are shown the STFT (in gray) and the frequency responses of a 12thorder model (heavy dark line) and a 40thorder model (thin dark line). This is in contrast to homomorphic Fig.6 Effect of predictor order on: (a) estimating the spectral envelope of a 4kHz bandwidth signal.6 offers a suggestion.5 Comparison of shorttime Fourier analysis with linear predictive analysis.
and valleys of the STFT. see Figure 5. In a particular application. Note the sharp decrease from p = 0 (no prediction implies V (0) = 1) to p = 1 and the less abrupt decrease thereafter. In this case. that H(ej2πf/fs)2 → Sˆn(ej2πf/fs)2.1) above. is comprised . Figure 6. notice that the meansquared error curve flattens out above about p = 12 and then decreases modestly thereafter. the radiation filtering is highpass.8. the LPC spectrum with p = 12. the vocal tract model is highly influenced by the pitch harmonics in the shorttime spectrum. it follows that the glottal pulse spectrum is lowpass. a homomorphically smoothed spectrum using 13 cepstrum values. Observe that p = 40 gives a much different result. for adult speakers. As an example. From the acoustic theory of speech production. It can be shown that if p is increased to the length of the windowed speech segment. but as pointed out in (A. the prediction order is generally fixed at a value that captures the general spectral shape due to the glottal pulse.6(b) shows the normalized meansquared prediction error V (p) = E(p)/φ[0] as a function of p for the segment of speech used to produce Figure 6. and a melfrequency spectrum. The choice p = 12 gives a good match to the general shape of the STFT. and radiation. E(2M) does not go to zero [75]. vocal tract resonances. which shows a comparison between the shorttime Fourier spectrum. Furthermore. and the vocal tract imposes a resonance structure that.6(a). highlighting the formant structure imposed by the vocal tract filter while ignoring the periodic pitch structure.
For example. but is less problematic for male voices where the spacing between harmonics is generally much smaller. 6. the combination of the lowpass glottal pulse spectrum and the highpass filtering of radiation are usually adequately represented by one or two additional complex pole pairs.5 where the sampling rate was fs = 16000 Hz. particularly when they are used in speech coding. . 6. For the sampled speech signal. the speaker was a highpitched female. a variety of different equivalent representations can be obtained. for a sampling rate of fs = 8000 Hz. These different representations are important. This effect is a serious limitation for highpitched voices. it is common to use a predictor order of 10–12. In that example. this leads to a rule of thumb of p = 4 + fs/1000. Note that in Figure 6. a predictor order of p = 22 gave a good representation of the overall shape and resonance structure of the speech segment over the band from 0 to 8000 Hz. When coupled with an estimate of one resonance per kilohertz. and the wide spacing between harmonics causes the peaks of the linear predictive vocal tract model to be biased toward those harmonics.6 illustrates an important point about linear predictive analysis.5 Equivalent Representations The basic parameters obtained by linear predictive analysis are the gain G and the prediction coefficients {αk}.5 Equivalent Representations 89 The example of Figure 6. From these.of about one resonance per kilohertz of frequency [101].
6. (6.5) shows that the system function of the prediction error filter is a polynomial in z−1 and therefore it can be represented in terms of its zeros as A(z) = 1 − _p k=1 αkz −k = _p k=1 (1 − zkz −1).because we often wish to quantize the model parameters for efficiency in storage or transmission of coded speech. we give only a brief summary of the most important equivalent representations. then it can be expected that roughly fs/1000 of the roots will be close in frequency (angle in the zplane) to the formant frequencies. These are the poles of H(z) that model the formant resonances. if the model order is chosen judiciously as discussed in the previous section. Note that eight (four complex conjugate pairs) of the roots are close to the unit circle. Therefore.7 shows an example of the roots (marked with ×) of a 12thorder predictor. In this section. The remaining four roots lie well within the . Figure 6.26) According to (6. the zeros of A(z) are the poles of H(z).5.6).1 Roots of Prediction Error System Function Equation (6.
not often invoked. 90 Linear Predictive Analysis Fig.4(a) with an additional section with kp+1 = . from which it follows that P(z) and Q(z) are system functions of lattice filters obtained by extending Figure 6. it is important to maintain high accuracy in the location of the zeros of A(z).27a) Q(z) = A(z) − z −(p+1)A(z −1).2 LSP Coefficients A much more desirable alternative to quantization of the roots of A(z) was introduced by Itakura [52]. The prediction coefficients are perhaps the most susceptible to quantization errors of all the equivalent representations. It is wellknown that the roots of a polynomial are highly sensitive to errors in its coefficients — all the roots being a function of all the coefficients [89].7 Poles of H(z) (zeros of A(z)) marked with × and LSP roots marked with ∗ and o. One possibility. which means that they only provide for the overall spectral shaping resulting from the glottal and radiation spectral shaping.27b) 6 Note the similarity to (6.unit circle. (6.5. who defined the line spectrum pair (LSP) polynomials6 P(z) = A(z) + z −(p+1)A(z −1) (6. 6. would be to factor the polynomial as in (6.23).26) and then quantize each root (magnitude and angle) individually. 6. Since the pole locations are crucial to accurate representation of the spectrum.
28a) Q(z) = p_+1 k=1 (1 − ejΘkz −1). (6.∓1.1) All the roots of P(z) and Q(z) are on the unit circle. (LSP.4) The LSFs are close together when the roots of A(z) . i. the LSFs are interlaced over the range of angles 0 to π. then P(−1) = 0 and Q(1) = 0. An illustration of the LSP representation is given in Figure 6. This new representation has some very interesting and useful properties that are confirmed by Figure 6.. P(z). which shows the roots of A(z).7 and summarized below [52. Ωk and Θk. respectively.e.7. (LSP. are called the line spectrum frequencies or LSFs.28b) The normalized discretetime frequencies (angles in the zplane).5 Equivalent Representations 91 This transformation of the linear prediction parameters is invertible: to recover A(z) from P(z) and Q(z) simply add the two equations to obtain A(z) = (P(z) + Q(z))/2. 6. (LSP. 119]..3) The PARCOR coefficients corresponding to A(z) satisfy ki < 1 if and only if the roots of P(z) and Q(z) alternate on the unit circle. and Q(z) for a 12thorder predictor. they can be represented as P(z) = p_+1 k=1 (1 − ejΩkz −1) (6. i. Knowing the p LSFs that lie in the range 0 < ω < π is sufficient to completely define the LSP polynomials and therefore A(z). (LSP.2) If p is an even integer.e.
a more direct recursive computation is possible [101].16) gives the recursion for computingˆh[n] from the predictor coefficients and gain.14). (6.3. since H(z) is a minimumphase system (all its poles inside the unit circle). If property (LSP. However. h[n]. These properties make it possible to represent the linear predictor by quantized differences between the successive LSFs [119].5.17) gives the inverse recursion .3 Cepstrum of Vocal Tract Impulse Response One of the most useful alternative representations of the linear predictor is the cepstrum of the impulse response.29) or a closed form expression for the impulse response can be obtained by making a partial fraction expansion of the system function H(z). because of the minimum phase condition. 92 Linear Predictive Analysis 6. the reconstructed polynomial A(z) will still have its zeros inside the unit circle. The impulse response can be computed recursively through the difference equation: h[n] = _p k=1 αkh[n − k] + Gδ[n]. which would require that the poles of H(z) be found by polynomial rooting. Then.3. and (5. of the vocal tract filter. any desired number of cepstrum values can be computed. Equation (5.are close to the unit circle. it follows that the complex cepstrum of h[n] can be computed using (5.3) is maintained through the quantization. As discussed in Section 5.
The resulting algorithm is given below. we can simply use them at step (D. i − 1 α (i−1) j= α (i) . . p kp = α (p) p (P.2.2) of the Levinson–Durbin algorithm thereby obtaining an algorithm for converting PARCORs to predictor coefficients. 6.5 Equivalent Representations 93 PredictortoPARCOR Algorithm α (p) j = αj j = 1. .4 PARCOR Coefficients We have seen that the PARCOR coefficients are bounded by ±1. . . These relationships are particularly useful in speech recognition applications where a small number of cepstrum values is used as a feature vector. .1) for i = p. This makes them an attractive parameter for quantization. . . we can compute the PARCOR coefficients from a given set of predictor coefficients.2. .2 for j = 1. By working backward through the Levinson– Durbin algorithm.5. . . . We have mentioned that if we have a set of PARCOR coefficients.p − 1. .for computing the predictor coefficients from the complex cepstrum of the vocal tract impulse response. 6.
j + kiα (i) i−j 1 − k2 i (P.2) end ki−1 = α (i−1) i−1 (P.3) end 6.5.5 Log Area Coefficients As mentioned above, the lattice filter representation of the vocal tract system is strongly related to concatenated acoustic tube models of the physics of sound propagation in the vocal tract. Such models are characterized by a set of tube crosssectional areas denoted Ai. These “area function” parameters are useful alternative representations of the vocal tract model obtained from linear predictive analysis. Specifically, the log area ratio parameters are defined as gi = log _ Ai+1 Ai _ = log _ 1 − ki 1 + ki _ , (6.30)
where Ai and Ai+1 are the areas of two successive tubes and the PARCOR coefficient −ki is the reflection coefficient for sound waves impinging on the junction between the two tubes [3, 78, 101]. The inverse transformation (from gi to ki) is ki = 1 − egi 1 + egi . (6.31) The PARCOR coefficients can be converted to predictor coefficients if required using the technique discussed in Section 6.5.4. Viswanathan and Makhoul [131] showed that the frequency response of a vocal tract filter represented by quantized log area ratio coefficients is relatively insensitive to quantization errors. 94 Linear Predictive Analysis 6.6 The Role of Linear Prediction In this chapter, we have discussed many of the fundamentals of linear predictive analysis of speech. The value of linear predictive analysis stems from its ability to represent in a compact form the part of the speech model associated with the vocal tract, which in turn is closely related to the phonemic representation of the speech signal. Over 40 years of research on linear predictive methods have yielded a wealth of knowledge that has been widely and effectively applied in almost every area of speech processing, but especially in speech coding, speech synthesis,
and speech recognition. The present chapter and Chapters 3–5 provide the basis for the remaining chapters of this text, which focus on these particular application areas. 7 Digital Speech Coding This chapter, the first of three on digital speech processing applications, focuses on specific techniques that are used in digital speech coding. We begin by describing the basic operation of sampling a speech signal and directly quantizing and encoding of the samples. The remainder of the chapter discusses a wide variety of techniques that represent the speech signal in terms of parametric models of speech production and perception. 7.1 Sampling and Quantization of Speech (PCM) In any application of digital signal processing (DSP), the first step is sampling and quantization of the resulting samples into digital form. These operations, which comprise the process of AtoD conversion, are depicted for convenience in analysis and discussion in the block diagram of Figure 7.1.1 1 Current AtoD and DtoA converters use oversampling techniques to implement the functions depicted in Figure 7.1. 95 96 Digital Speech Coding Sampler
Quantizer Q{ } xc (t) x[n] xˆ[n] Encoder c[n] AtoD Converter Decoder c′[n] xˆ′[n] Δ TΔ Reconstruction Filter DtoA Converter ) ( ˆ t x′c Fig. 7.1 The operations of sampling and quantization. The sampler produces a sequence of numbers x[n] = xc(nT), where T is the sampling period and fs = 1/T is the sampling frequency. Theoretically, the samples x[n] are real numbers that are useful for theoretical analysis, but never available for computations. The well known Shannon sampling theorem states that a bandlimited signal can be reconstructed exactly from samples taken at twice the highest frequency in the input signal spectrum (typically between 7 and 20 kHz for speech and audio signals) [89]. Often, we use a lowpass filter to remove spectral information above some frequency of interest (e.g., 4 kHz) and then use a sampling rate such as 8000 samples/s to avoid aliasing distortion [89]. Figure 7.2 depicts a typical quantizer definition. The quantizer simply takes the real number inputs x[n] and assigns an output ˆx[n] according
to the nonlinear discreteoutput mapping Q{ }.2 In the case of the example of Figure 7.2, the output samples are mapped to one of eight possible values, with samples within the peaktopeak range being rounded and samples outside the range being “clipped” to either the maximum positive or negative level. For samples within range, the quantization error, defined as e[n] = ˆx[n] − x[n] (7.1) 2 Note that in this chapter, the notation ˆx[n] denotes the quantized version of x[n], not the complex cepstrum of x[n] as in Chapter 5. 7.1 Sampling and Quantization of Speech (PCM) 97 Fig. 7.2 8level midtread quantizer. satisfies the condition −Δ/2 < e[n] ≤ Δ/2, (7.2) where Δ is the quantizer step size. A Bbit quantizer such as the one shown in Figure 7.2 has 2B levels (1 bit usually signals the sign). Therefore, if the peaktopeak range is 2Xm, the step size will be Δ = 2Xm/2B. Because the quantization levels are uniformly spaced by Δ, such quantizers are called uniform quantizers. Traditionally, representation of a speech signal by binarycoded quantized samples is called pulsecode modulation (or just PCM) because binary numbers can be represented for transmission as on/off pulse amplitude modulation. The block marked “encoder” in Figure 7.1 represents the assigning of a binary code word to each quantization level. These code words represent the quantized signal amplitudes, and generally, as in Figure 7.2, these code words are chosen to correspond to some convenient binary
Most signals. i.3 In cases where accurate amplitude calibration is desired. However. This means that we need a large peaktopeak range so as to avoid clipping of the loudest sounds.1. their amplitudes vary greatly between voiced and unvoiced sounds and between speakers. speech especially.4 Generally. (7.2) states that the size of the maximum quantization error grows. but this is not the case if the step size is adapted from sampletosample. For a given peaktopeak range for a quantizer such as the one in Figure 7. The operation of quantization confronts us with a dilemma. for a given number of levels (bits). Therefore. Furthermore. . Note that the combination of the decoder and lowpass reconstruction filter represents the operation of DtoA conversion [89]. we do not need to be concerned with the fine distinction between the quantized samples and the coded samples when we have a fixed step size. as the peaktopeak range increases. the step size Δ = 2Xm/2B is proportional to the peaktopeak range.e. the maximum size of the quantization error is the same whether the signal sample is large or small.number system such that arithmetic can be done on the code words as 98 Digital Speech Coding if they were proportional to the signal samples.. have a wide dynamic range. the step size could be applied as depicted in the block labeled “Decoder” at the bottom of Figure 7.2 we can decrease the quantization error only by adding more levels (bits). for a uniform quantizer.
quantizer. 4 The _ denotes the possibility of errors in the codewords.5 This value is more than adequate and much more than desired for most speech communication applications. and only the binary code words c[n] are available.1 Sampling and Quantization of Speech (PCM) 99 techniques for significantly reducing the bit rate while maintaining an adequate level of speech quality. singing. 5 Even higher rates (24 bits and 96 kHz sampling rate) are used for highquality audio. The standard values for sampling and quantizing sound signals (speech. instrumental music) are B = 16 and fs = 44.This is the fundamental problem of quantization. 7. the quantization error sequence. and coder are all integrated into one system.1. The data rate (measured in bits/second) of sampled and quantized speech signals is I = B · fs. both these simple approaches degrade the perceived quality of the speech signal. This chapter deals with a wide range of 3 In a real AtoD converter. 7. Symbol errors would cause additional error in the reconstructed samples. If the number of bits in the quantizer is reasonably high and no clipping occurs. although it is completely . however. the sampler.600 bits/s.1 or 48 kHz. The bit rate can be lowered by using fewer bits/sample and/or using a lower sampling rate. This leads to a bitrate of I = 16 · 44100 = 705.1 Uniform Quantization Noise Analysis A more quantitative description of the effect of quantization can be obtained using random signal analysis applied to the quantization error.
The condition “uncorrelated” . notably smooth input probability density functions and high rate quantizers. i. Under these conditions.determined by the signal amplitudes.3) The amplitudes of noise samples are uniformly distributed across the range −Δ/2 < e[n] ≤ Δ/2.e. nevertheless behaves as if it is a random signal with the following properties [12]: (Q.. resulting in average power σ2 e =Δ2/12. (Q. With these assumptions. These results can be easily shown to hold in the simple case of a uniformly distributed memoryless input and Bennett has shown how the result can be extended to inputs with smooth densities if the bit rate is assumed high [12].1) The noise samples appear to be6 uncorrelated with the signal samples.2) Under certain assumptions. the noise samples appear to be uncorrelated from sampletosample. it can be shown that if the output levels of the quantizer are optimized. as commonly stated). Bennett has shown that the correlations are only small because the error is small and. then the quantizer error will be uncorrelated with the quantizer output (however not the quantizer input. (Q. it is possible to derive the following formula for the signaltoquantizingnoise ratio (in dB) of a Bbit 6 By “appear to be” we mean that measured correlations are small. e[n] acts like a white noise sequence. These simplifying assumptions allow a linear analysis that yields accurate results if the signal is not too coarsely quantized. the correlations are equal to the negative of the error variance [12]. under suitable conditions.
if the bit rate is not large [12].3) is an increasingly good approximation as the bit rate (or the number of quantization levels) gets large. and 10 bits.3) is a reasonable estimate of SNR.3) and the dark dashed lines are measured values for uniform quantization.02B + 4. The faint dashed lines are from (7. however. but in this case the error is a deterministic function of the input and hence it cannot be independent of the input.78 − 20log10 _ Xm σx _ . 100 Digital Speech Coding uniform quantizer: SNRQ = 10log _ σ2x σ2 e _ = 6. Note that Xm is a fixed parameter of the quantizer.often implies independence.3) where σx and σe are the rms values of the input signal and quantization noise samples. The measurements were done by quantizing 16bit samples to 8. There is good agreement between these graphs indicating that (7. The formula of Equation (7.3 shows a comparison of (7.3) with signaltoquantizationnoise ratios measured for speech signals. respectively. Figure 7. It can be way off. (7. while σx . 9.
On the other hand.9.3)—light dashed lines. When σx gets close to Xm. In other words. This accounts for the precipitous fall in SNR for 1<Xm/σx < 8.depends on the input signal level. the ratio Xm/σx decreases moving to the left in Figure 7. 10: Equation (7.3 Comparison of μlaw and linear quantization for B = 8. Measured uniform quantization SNR — dark dashed lines. 0 10 20 30 40 50 60 8 9 10 Comparison of Law and Uniform Quantizers SNR in dB Fig. 7.1 Sampling and Quantization of Speech (PCM) 101 Also. and the assumptions underlying (7. As signal level increases.3) and Figure 7. it is also important to note that halving σx decreases the SNR by 6 dB. 7. increasing B by 1 bit (doubling the number of quantization levels) increases the SNR by 6 dB. many samples are clipped.3 show that with all other parameters being fixed. it should be noted that (7. μlaw (μ = 100) compression — solid lines. cutting the signal level in half is like throwing .3.3) no longer hold.
the SNR would be constant regardless of the size of the signal. μlaw compression (of the dynamic range) of the signal.away one bit (half of the levels) of the quantizer. 7. If μ is large (> 100) this nonlinear transformation has the effect of distributing the effective quantization levels uniformly for small samples and logarithmically over .3 decrease linearly with increasing values of log10(Xm/σx). Quantization of μlaw compressed signals is often termed logPCM. it is exceedingly important to keep input signal levels as high as possible without clipping. If the quantization error were proportional to the signal amplitude.2 μLaw Quantization Also note that the SNR curves in Figure 7. This is because the size of the quantization noise remains the same as the signal level decreases. As a compromise. but this is not possible because the log function is illbehaved for small values of x[n]. defined as y[n] = Xm log ! 1+μ x[n] Xm " log(1 + μ) · sign(x[n]) (7. Thus.4) can be used prior to uniform quantization [118].1. This could be achieved if log(x[n]) were quantized instead of x[n].
1.” If the digital output of a codec is used as input to a speech processing algorithm. This is an example of where speech quality is deliberately compromised in order to achieve a much lower bit rate. It is .3 show that the signaltonoise ratio with μlaw compression remains relatively constant over a wide range of input levels. In particular. it was perceived to render speech equivalent to speech transmitted over the best long distance lines. μlaw quantization for speech is standardized in the CCITT G. 8bit.4). logPCM with fs = 8000 samples/s is widely used for digital telephony. Such an AtoD/DtoA device is called a encoder/decoder or “codec. readily available hardware devices convert analog signals directly into binarycoded μlaw samples and also expand compressed samples back to uniform scale for conversion back to analog signals. 8bit logPCM introduces little or no perceptible distortion. Once the bandwidth restriction has been imposed.711 standard.3 NonUniform and Adaptive Quantization μlaw compression is an example of nonuniform quantization. it is generally necessary to restore the linearity of the amplitudes through the inverse of (7. 7. μ = 255. Nowadays. The flat curves in Figure 7.the remaining range of the input samples. 102 Digital Speech Coding This configuration is often referred to as “toll quality” because when it was introduced at the beginning of the digital telephony era.
To do this analytically. however. A more rigorous approach is to design a nonuniform quantizer that minimizes the meansquared quantization error. To apply this idea to the design of a nonuniform quantizer for speech requires an assumption of an analytical form for the probability distribution or some algorithmic approach based on measured distributions. 7 Lloyd’s work was initially published in a Bell Laboratories Technical Note with portions of the material having been presented at the Institute of Mathematical Statistics Meeting in Atlantic City. it is necessary to know the probability distribution of the signal sample values so that the most probable samples. New Jersey in September 1957. Lloyd [71] gave an algorithm for designing optimum nonuniform quantizers based on sampled speech signals. Paez and Glisson [91] gave an algorithm for designing optimum quantizers for assumed Laplace and Gamma probability densities. will incur less error than the least probable samples.7 Optimum nonuniform quantizers can improve the signaltonoise ratio by as much as 6 dB over μlaw quantizers with the same number of bits. which are useful approximations to measured distributions for speech. which for speech are the low amplitude samples.based on the intuitive notion of constant percentage error. The fundamentals of optimum quantization were established by Lloyd [71] and Max [79]. Subsequently this pioneering work was . but with little or no improvement in perceived quality of reproduction.
With this definition. To amortize this overhead. for example. On the other hand. Such adaption of the step size is equivalent to fixed quantization of the signal after division by the rms signal amplitude (Eˆn)1/2.6). (7. the quantizer is a feedback adaptive quantizer. the step size could satisfy Δ[n] = Δ0(Eˆn)1/2. 7. thereby adding to the information rate. The adaptation speed can be adjusted by varying the analysis window size. feedforward quantization is usually applied to blocks of speech samples. if the step size control is based on past quantized samples. For example. Jayant [57] studied a class of feedback adaptive quantizers where the . the quantizer is called a feedforward adaptive quantizer. the step size will go up and down with the local rms amplitude of the speech signal. When the quantizer in a PCM system is adaptive. as in (4. If the step size control is based on the unquantized signal samples.published in the open literature in March 1982 [71].5) where Δ0 is a constant and Eˆn is the shorttime energy of the speech signal defined.2 Digital Speech Coding 103 Another way to deal with the wide dynamic range of speech is to let the quantizer step size vary with time. the system is called an adaptive PCM (APCM) system. and the step size information must be transmitted or stored along with the quantized sample.
if the previous sample is quantized using step size Δ[n − 1] to one of the low quantization levels. 58]. engineers and speech scientists have worked toward . the complexity (in terms of digital computation) is often of equal concern. In addition to the two dimensions of quality and bit rate. In this approach. The main objective in digital speech coding (sometimes called speech compression) is to lower the bit rate while maintaining an adequate level of perceptual fidelity. Reducing fs requires bandwidth reduction. and reducing B too much will introduce audible distortion that may resemble random noise. Since the 1930s. but perceptual quality may suffer. By basing the adaptation on the previous sample.8 By reducing B or fs.step size Δ[n] is a function of the step size for the previous sample. it is not necessary to transmit the step size. Adaptive quantizers can achieve about 6 dB improvement (equivalent to adding one bit) over fixed quantizers [57. the bit rate can be reduced. the step size is increased if the previous sample was quantized in one of the highest levels. On the other hand. 7. Virtually perfect perceptual quality is achievable with high data 104 Digital Speech Coding rate.2 Digital Speech Coding As we have seen. Δ[n − 1]. the data rate of sampled and quantized speech is B · fs. It can be derived at the decoder by the same algorithm as used for encoding. then the step size is decreased for the next sample.
Nonuniform quantization and adaptive quantization are simple attempts to build in properties of the speech signal (timevarying amplitude distribution) and speech perception (quantization noise is masked by loud sounds). Most modern modelbased coding methods are based on the notion that the speech signal can be represented in terms of an excitation signal and a timevarying vocal tract system. and the latter modelbased coders. Straightforward sampling and uniform quantization (PCM) is perhaps the only pure waveform coder.the ultimate goal of achieving more efficient representations. This work has led to a basic approach in speech coding that is based on the use ofDSP techniques and the incorporation of knowledge of speech production and perception into the quantization process. Traditionally. There are many ways that this can be done. but these extensions still are aimed at preserving the waveform. Modelbased coders are designed for obtaining efficient digital representations of speech and only speech. and then a “quantized” speech signal can be reconstructed . These two components are quantized. speech coding methods have been classified according to whether they attempt to preserve the waveform of the speech signal or whether they only seek to maintain an acceptable level of perceptual quality and intelligibility. The former are generally called waveform coders.
which means that for a fixed number of bits.4 is a factor of five lower than that of the signal itself.4 illustrates why linear predictive analysis can be fruitfully applied for modelbased coding. Figure 7.by exciting the quantized filter with the quantized excitation. Note that the prediction error sequence amplitude of Figure 7.2 Speech Waveform Prediction Error Sequence sample index Fig.4 Speech signal and corresponding prediction error signal for a 12thorder predictor.2 Digital Speech Coding 105 0 100 200 300 400 −1 0 1 0 100 200 300 400 −0. 7.2 0 0. the digital representation is often referred to as being “transparent. The upper plot is a segment of a speech signal. this segment of the prediction error could be quantized with a smaller step size than the waveform .” 7. The lower plot is the output of a linear prediction error filter A(z) (with p = 12) that was derived from the given segment by the techniques of Chapter 6. 8 If the signal that is reconstructed from the digital representation is not perceptibly different from the original analog speech signal.
Dudley early in the history of speech processing research . 106 Digital Speech Coding These coders.9 While direct implementation of this idea does not lead to a practical method of speech coding. are also called vocoders (voice coder) since they are based on the principles established by H. A more useful classification of speech coders focusses on how the speech production and perception models are incorporated into the quantization process. Today. the line between waveform coders and modelbased coders is not distinct. an approximation to the original segment of speech would be obtained. it is nevertheless suggestive of more practical schemes which we will discuss.itself. 9To implement this timevarying inverse filtering/reconstruction requires care in fitting the segments of the signals together at the block boundaries. Overlapadd methods [89] can be used effectively for this purpose. If the quantized prediction error (residual) segment is used as input to the corresponding vocal tract filter H(z) = 1/A(z). One class of systems simply attempts to extract an excitation signal and a vocal tract system from the speech signal without any attempt to preserve a relationship between the waveforms of the original and the quantized speech. which we designate as openloop coders. Such systems are attractive because they can be implemented with modest computation.
These compare the quantized output to the original input and attempt to minimize the difference between the two in some prescribed sense. Closed loop systems. These systems are generally classified as waveform coders. Figure 7. since they explicitly attempt to minimize a timedomain distortion measure. Differential PCM systems are simple examples of this class.1 Predictive Coding The essential features of predictive coding of speech were set forth in a classic paper by Atal and Schroeder [5]. often do a good job of preserving the speech waveform while employing many of the same techniques used in openloop systems. although the basic principles of predictive quantization were introduced by Cutler [26]. A second class of coders employs the source/system model for speech production inside a feedback loop.3 ClosedLoop Coders 7. modelbased systems that also preserve the .3. 7. but increased availability of computational power has made it possible to implement much more sophisticated closedloop systems called analysisbysynthesis coders.[30]. but we prefer to emphasize that they are closedloop. and thus are called closedloop coders.5 shows a general block diagram of a large class of speech coding systems that are called adaptive differential PCM (ADPCM) systems.
. but irrespective of the type of quantizer. The reader should ignore initially the blocks concerned with adaptation and all the dotted control lines and focus instead on the core DPCM system. The block labeled P is a linear predictor. as shown in Figure 7. the signal ˜x[n] is predicted based on p past samples of the signal ˆx[n].10 The signal d[n] = x[n] − ˜x[n]. The quantizer Q can have a number of levels ranging from 2 to much higher. the predictor parameters are estimated from the input signal x[n].waveform.5 General block diagram for adaptive differential PCM (ADPCM). The latter would be termed a feedback adaptive predictor. although. the 10 In a feedforward adaptive predictor. they can also be estimated from past samples of the reconstructed signal ˆx[n]. and it can be adaptive or not. i. the difference between the input and the predicted signal.e.6) 7. where e[n] is the quantization error.5. Finally. 108 Digital Speech Coding . (7. it can be uniform or nonuniform. the quantized output can be expressed as ˆ d[n] = d[n] + e[n]. 7. so ˜x[n] = _p k=1 αk ˆx[n − k]. is the input to the quantizer. which is comprised of a feedback structure that includes the blocks labeled Q and P.3 ClosedLoop Coders 107 Fig.
8) This is the key result for DPCM systems. then the variance of d[n] will be less than the variance of x[n] so it will be possible to use a smaller step size and therefore to reconstruct ˆx[n] with less error than if x[n] were quantized directly. (7. since ˆx[n] = ˜x[n] + ˆ d[n] it follows that ˆx[n] = x[n] + e[n].5 contains within it a source/system speech model. which. is the system needed to reconstruct the quantized speech signal ˆx[n] from the coded difference signal input. The feedback structure of Figure 7.relationship between ˆx[n] and ˆ d[n] is ˆx[n] = _p k=1 αk ˆx[n − k] + ˆ d[n]. the signal ˆx[n] is the output of what we have called the “vocal tract filter. Finally.e. A simple. but informative modification leads to SNR = σ2x σ2 d · σ2 d .” and ˆ d[n] is the quantized excitation signal. If the prediction is good..5(b). (7.8).5. the quantization error in ˆx[n] is identical to the quantization error in the quantized excitation signal ˆ d[n]. the signaltonoise ratio of the ADPCM system is SNR = σ2x /σ2 e . From (7. as shown in Figure 7. It says that no matter what predictor is used in Figure 7.7) i.
e. is given by (7. which.3 ClosedLoop Coders 109 the correlation between the signal and the quantization error.σ2 e = GP · SNRQ. As indicated.9) where the obvious definitions apply.5.e. step size information must be part of the digital representation. As shown in Figure 7. for fine quantization (i. the quantizer can be either feedforward. it can be shown that GP = σ2x σ2 d = 1 1− _p k=1 αkρ[k] . The quantity GP is called the prediction gain. The number of bits B determines the bitrate of the coded difference signal.. SNRQ is the signaltonoise ratio of the quantizer. (7..3). if the quantizer is feedforward adaptive. large number of quantization levels with small step size). Neglecting 7. i. if GP > 1 it represents an improvement gained by placing a predictor based on the speech model inside the feedback loop around the quantizer.10) . (7.or feedbackadaptive.
5 1 predictor order p autocorrelation lag Fig.10) as a function of predictor order p. Note that a firstorder fixed predictor can yield about 6 dB prediction gain so that either the quantizer can have 1 bit less or the 0 2 4 6 8 10 12 0 5 10 G p in dB 0 2 4 6 8 10 12 −0. Fixed predictors can be designed based on longterm average correlation functions.where ρ[k] = φ[k]/φ[0] is the normalized autocorrelation function used to compute the optimum predictor coefficients. 7.6 Longterm autocorrelation function ρ[m] (lower plot) and corresponding prediction gain GP .5 0 0.6 shows an estimated longterm autocorrelation function for speech sampled at an 8 kHz sampling rate. The upper plot shows 10log10GP computed from (7. The lower figure in Figure 7. . and adaptive predictors can be either feedforward or feedback (derived by analyzing past samples of the quantized speech signal reconstructed as part of the coder). The predictor can be either fixed or adaptive.
110 Digital Speech Coding reconstructed signal can have a 6 dB higher SNR. DM systems originated in the classic work of Cutler [26] and deJager [28]. To see how such systems can work. (7. Here we can only mention a few of the most important.10) it follows that the prediction gain for a firstorder predictor is GP = 1 1 − ρ2[1] . many systems based upon the basic principle of differential quantization have been proposed.5.3. and adapting the predictor at a phoneme rate can produce an additional 4 dB of gain [84]. and implemented as standard systems. they nevertheless fit neatly into the general theory of predictive coding. note the nature of the longterm autocorrelation function . so from (7. While DM systems evolved somewhat independently of the more general differential coding methods. 7.2 Delta Modulation Delta modulation (DM) systems are the simplest differential coding systems since they use a 1bit quantizer and usually only a firstorder predictor. and not surprisingly. Great flexibility is inherent in the block diagram of Figure 7. note that the optimum predictor coefficient for a firstorder predictor is α1 = ρ[1]. studied. Higherorder fixed predictors can achieve about 4 dB more prediction gain.11) Furthermore.
as fs gets large. For example. Delta modulation systems can be implemented with very simple hardware. i.in Figure 7. we could expect that the correlation value ρ[1] would lie on a smooth curve interpolating between samples 0 and 1 in Figure 7. Jayant [56] showed that for bandlimited (3 kHz) speech.. 56]. and in contrast to more general predictive coding systems. this can be compensated by the prediction gain due to oversampling.3 ClosedLoop Coders 111 sampling rate with a fixed predictor.6. both ρ[1] and α1 approach unity and GP gets large. they are not generally implemented with block processing.6. This limitation can be mitigated with only modest increase in complexity by using an adaptive 1bit quantizer [47. Instead. If the same bandwidth speech signal were oversampled at a sampling rate higher than 8000 samples/s. Thus. being equal to fs. adaptation algorithms are usually based on the bit stream at the output of the 1bit quantizer. even though a 1bit quantizer has a very low SNR. must be very high to achieve good quality reproduction. Their bit rate. This correlation function is for fs = 8000 samples/s. adaptive delta modulation (ADM) with bit rate (sampling rate) 40 kbits/s has about the same SNR as 6bit .e. The simplest delta modulators use a very high 7.
24. For these reasons.3. 32.log PCM sampled at 6.6 kHz. 7. Generally. only bit synchronization is required in transmission. (3) Adding firstorder fixed or adaptive prediction improved the SNR by about 4 dB over APCM. he showed that doubling the sampling rate of his ADM system improved the SNR by about 10 dB. fixed DPCM. which introduces delay. 4. Adaptive delta modulation has the following advantages: it is very simple to implement. Noll compared log PCM. For all bit rates (16. and three ADPCM systems all using quantizers of 2. and 40 kbps).5 are implemented on short blocks of input signal samples. Furthermore. adaptive differential PCM systems can have any combination of adaptive or fixed quantizers and/or predictors. adaptive delta modulation (ADM) is still used for terminalcostsensitive digital transmission applications. (2) Adapting the quantizer (APCM) improved SNR by 6 dB. it has very low coding delay.3 Adaptive Differential PCM Systems Differential quantization with a multibit quantizer is called differential PCM (DPCM). . and 5bits with 8 kHz sampling rate. adaptive PCM. 3. As depicted in Figure 7. and it can be adjusted to be robust to channel errors. the results were as follows11: (1) Log PCM had the lowest SNR.5. the operations depicted in Figure 7. A careful comparison of a variety of ADPCM systems by Noll [84] gives a valuable perspective on the relative contributions of the quantizer and predictor.
G. 112 Digital Speech Coding (4) A fourthorder adaptive predictor added about 4 dB. The classic paper by Atal and Schroeder [5] contained a number of ideas that have since been applied with great success in adaptive predictive coding of speech. it can be filtered with the inverse system to restore the spectral balance. which we have seen is simply added to the input signal in the reconstruction process. See Section 7. One of these concerns the quantization noise introduced by ADPCM. G. filter the speech with a linear filter that boosts the highfrequency (HF) part of the spectrum.727). After reconstruction of the preemphasized speech. and in the process. and increasing the predictor order to 12 added only 2 dB more.3.3.e..721.1 for a discussion of this issue. which are often operated at 32 kbps where quality is superior to logPCM (ITU G. the noise spectrum will take on the shape of the deemphasis . it is clear that the noise will be most prominent in the speech at high frequencies where speech spectrum amplitudes are low.726.11 The bit rates mentioned do not include overhead information for the quantized prediction coefficients and quantizer step sizes. A simple solution to this problem is to preemphasize the speech signal before coding. i.711) at 64 kbps. The superior performance of ADPCM relative to logPCM and its relatively low computational demands have led to several standard versions (ITU G. If we invoke the white noise approximation for quantization error.
Including the simplest longdelay predictor would change (7. Another idea that has been applied effectively in ADPCM systems as well as analysisbysynthesis systems is to include a longdelay predictor to capture the periodicity as well as the shortdelay correlation 12Values around γ = 0. A simple preemphasis system of this type has system function (1 − γz−1) where γ is less than one.12 A more sophisticated application of this idea is to replace the simple fixed preemphasis filter with a timevarying filter designed to shape the quantization noise spectrum.4 for fs = 8 kHz work best. All these approaches raise the noise spectrum at low frequencies and lower it at high frequencies. 7.filter spectrum. the low frequency noise is emphasized too much.3 ClosedLoop Coders 113 inherent in voiced speech. If γ is too close to one.6) to ˜x[n] = βˆx[n − M] + _p k=1 αk(ˆx[n] − βˆx[n − M]). An equivalent approach is to define a new feedback coding system where the quantization noise is computed as the difference between the input and output of the quantizer and then shaped before adding to the prediction residual. (7. thus taking advantage of the masking effects of the prominent low frequencies in speech [2].12) The parameter M is essentially the pitch period (in samples) of voiced .
Then the shortdelay predictor parameters αk are estimated from the output of the longdelay prediction error filter. (7. . Even more complicated predictors with multiple delays and gains have been found to improve the performance of ADPCM systems significantly [2].13) Adding longdelay prediction takes more information out of the speech signal and encodes it in the predictor. One suboptimal approach is to estimate β and M first by determining M that maximizes the correlation around the expected pitch period and setting β equal to the normalized correlation at M. the “vocal tract system” used in the decoder would have system function H(z) = 1 1− _p k=1 αkz −k _ 1 1 − βz−M _ . The predictor parameters in (7.12) cannot be jointly optimized.speech.13 When this type of predictor is used. and the parameter β accounts for amplitude variations between periods.
7. almost all of which are preferable to direct quantization of the predictor coefficients themselves.g. 50–100 times/s. the predictor coefficients and quantizer step size must be included as auxiliary data (side information) along with the quantized difference 13 Normally. this analysis would be performed on the input speech signal with the resulting predictor coefficients applied to the reconstructed signal as in (7.. . e.3. The PARCOR coefficients can be quantized with a nonuniform quantizer or transformed with an inverse sine or hyperbolic tangent function to flatten their statistical distribution and then quantized with a fixed uniform quantizer. The difference signal will have the same sampling rate as the input signal.5 shows that if the quantizer and predictor are fixed or feedbackadaptive. Otherwise. Each coefficient can be allocated a number of bits contingent on its importance in accurately representing the speech spectrum [131].1 Coding the ADPCM Parameters A glance at Figure 7.12). then only the coded difference signal is needed to reconstruct the speech signal (assuming that the decoder has the same coefficients and feedbackadaptation algorithms). 114 Digital Speech Coding signal.3. The step size and predictor parameters will be estimated and changed at a much lower sampling rate. The total bit rate will be the sum of all the bit rates. As discussed in Chapter 6. Predictor Quantization The predictor parameters must be quantized for efficient digital representation of the speech signal. the shortdelay predictors can be represented in many equivalent ways.
which requires 7 bits to cover the range of pitch periods to be expected with 8 kHz sampling rate of the input.7. The number of bits per coefficient ranged from 5 for each of the lowest two PARCORs down to 1 each for the six highestorder PARCORs. then a total of 4000 bps for the shortdelay predictor information is required.M. it is possible to update the shortdelay predictor 50 times/s for a total bit rate contribution of 2000 bps [2]. VQ can be applied in any context where a set of parameters naturally groups together into a vector (in the sense of 7. If the parameters are updated 100 times/s. Vector Quantization for Predictor Quantization Another approach to quantizing the predictor parameters is called vector quantization. To save bit rate. yielding a total of 40 bits per frame. or VQ [45]. This gave a total of 20 bits/frame and added 2000 bps to the overall bit rate. The basic principle of vector quantization is depicted in Figure 7. The long delay predictor has a delay parameter M.3 ClosedLoop Coders 115 Distortion Measure Code Book .Atal [2] reported a system which used a 20thorder predictor and quantized the resulting set of PARCOR coefficients after transformation with an inverse sine function. The longdelay predictor in the system mentioned above used delays of M − 1.M + 1 and three gains each quantized to 4 or 5 bit accuracy.
PARCOR coefficients. 7. The “training phase” is computationally intense. The design of a vector quantizer requires a training set consisting of a large number (L) of examples of vectors that are drawn from the same distribution as vectors to be quantized later. In VQ. v. the index i can be represented by a B bit number. cepstrum values. to be quantized is compared exhaustively to a “codebook” populated with representative vectors of that type. 2B codebook vectors are formed from the training set of vectors. is returned from the exhaustive search. a vector. This index i then represents the quantized vector in the sense that if the codebook is known. but need only be done . or line spectrum frequencies. The index i of the closest vector ˆvi to v.2B Vectors Table Lookup Code Book 2B Vectors vii vˆ k vˆ k vˆ i Encoder Decoder Fig. according to a prescribed distortion measure.7 Vector quantization. If the codebook has 2B entries. along with a distortion measure for comparing two vectors. the corresponding quantized vector ˆvi can be looked up. log area ratios. linear algebra).14 For example the predictor coefficients can be represented by vectors of predictor coefficients. Using an iterative process.
One advantage of VQ is that the distortion measure can be designed to sort the vectors into perceptually relevant prototype vectors.3.once. In order to reduce the bit rate for coding the difference samples. The resulting codebook is then used to quantize general test vectors. 116 Digital Speech Coding The quantization process is also time consuming because a given vector must be systematically compared with all the codebook vectors to determine which one it is closest to. it is desirable to use as few bits as possible for the quantizer. Difference Signal Quantization It is necessary to code the difference signal in ADPCM. some sort of block .. This has led to numerous innovations in codebook structure that speed up the look up process [45].3. If straightforward Bbit uniform quantization is used. This was the approach used in earlier studies by Noll [84] as mentioned in Section 7. and it is facilitated by the LBG algorithm [70].e. Since the difference signal has the same sampling rate as the input signal. L ≥ 10 · 2B. the additional bit rate would be B · fs. i. Typically a VQ training set consists of at least 10 times the ultimate number of codebook vectors. 14 Independent quantization of individual speech samples or individual predictor coefficients is called scalar quantization.
” which is a system that sets to zero value all samples whose amplitudes are within a threshold band and passes those samples above threshold unchanged.coding must be applied. this coding method needs a total of 5600 bps for the difference signal. The blocks of zero samples can be encoded efficiently using variablelengthtoblock codes yielding average bits/sample on the order of 0. 7.7 [2]. Atal and Schroeder [2.3. long runs of zero samples result. and the entropy of the quantized difference signal becomes less than one. 7] proposed to precede the quantizer by a “center clipper. Bit Rate for ADPCM Coders ADPCM systems normally operate with an 8 kHz sampling rate. the speech signal must be bandlimited to somewhat less than 4 kHz by a . When this bit rate of the differential coder is added to the bit rate for quantizing the predictor information (≈4000 bps) the total comes to approximately 10.3.000 bps. With this sampling rate. which is a compromise aimed at providing adequate intelligibility and moderate bit rate for voice communication. For a sampling rate of 8 kHz. the number of zero samples can be controlled.2 Quality vs. For high thresholds. One approach is to design a multibit quantizer that only operates on the largest samples of the difference signal. By adjusting the threshold.
quality degrades significantly. Adaptive prediction and quantization can easily lower the bit rate to 32 kbps with no degradation in perceived quality (with 7. its performance is fundamentally constrained by the fact that the difference signal has the same sampling rate as the input signal. and where adequate transmission and/or storage capacity is available to support bit rates on the order of 10 kbps. However.3. However. This increases the system complexity.lowpass filter. raising the bit rate for ADPCM above 64 kbps cannot improve the quality over logPCM because of the inherent frequency limitation. In order to achieve neartollquality at low bit rates. all the techniques discussed above and more must be brought into play. this approach is clearly not optimal since the center . Below this value. Thus. although not unreasonably so for current DSP hardware. Of course. ADPCM is an attractive coding method when toll quality is required at modest cost. 7.3 ClosedLoop Coders 117 64 kbps log PCM toll quality as a reference). the bit rate can be lowered below 32 kbps with only modest distortion until about 10 kbps.4 AnalysisbySynthesis Coding While ADPCM coding can produce excellent results at moderate bit rates. The center clipping quantizer produces a difference signal that can be coded efficiently.
What is needed is a way of creating an excitation signal for the vocal tract filter that is both efficient to code and also produces decoded speech of high quality. but with some significant modifications.8 are carried out on blocks of speech samples. ˆx[n]. and the output of the vocal tract filter. x[n]. d[n]. In particular.3. 7.clipper throws away information in order to obtain a sparse sequence. As in the case of the most sophisticated ADPCM systems. the difference. is filtered with a linear filter called the 118 Digital Speech Coding Perceptual Weighting W(z) Excitation Generator Vocal Tract Filter H(z) + .1 Basic AnalysisbySynthesis Coding System Figure 7. the operations of Figure 7. These systems are called “analysisbysynthesis coders” because the excitation is built up using an iterative process to produce a “synthetic” vocal tract filter output ˆx[n] that matches the input speech signal according to a perceptually weighted error criterion.8 shows the block diagram of another class of closedloop digital speech coders.4. This can be done within the same closedloop framework as ADPCM. between the input.
8 to the core ADPCM diagram of Figure 7.5. the reconstruction error is −d[n]. perceptual weighting filter.. where an adaptive quantization algorithm operates on d[n] to produce a quantized difference signal ˆ d[n]. In ADPCM. In analysisbysynthesis. ˆx[n] = x[n] − d[n]. but instead of the synthetic output ˆx[n].e. As the first step in coding a block of speech samples. i. W(z). the vocal tract model is in the same position in the closedloop system.8 Structure of analysisbysynthesis speech coders. and a perceptually . This is a key difference. a signal ˜x[n] predicted from ˆx[n] is subtracted from the input to form the difference signal. which is the input to the vocal tract system. In ADPCM. the excitation signal is determined from the perceptually weighted difference signal d_[n] by an algorithm represented by the block labeled “Excitation Generator. The perceptual weighting and excitation generator inside the dotted box play the role played by the quantizer in ADPCM.x[n] d[n] dˆ[n] xˆ[n] + d′[n] Fig. the synthetic output differs from the input x[n] by the quantization error. Then.” Note the similarity of Figure 7. both the vocal tract filter and the perceptual weighting filter are derived from a linear predictive analysis of the block. 7.
If α1 > α2 the frequency response is like a “controlled” inverse filter for H(z). where typical values of α1 = 0. and the zeros of W(z) are at the same angles but at radii of α1 times the radii of the poles of H(z).weighted version of that error is minimized in the meansquared sense by the selection of the excitation ˆ d[n]. d_[n] = d[n] ∗ w[n] with the weighting filter usually defined in terms of the vocal tract filter as the linear system with system function W(z) = A(z/α1) A(z/α2) = H(z/α2) H(z/α1) . this filter tends to emphasize the high frequencies (where the vocal tract . it is desirable to shape its spectrum to take advantage of perceptual masking effects.3 ClosedLoop Coders 119 to the vocal tract filter is determined so as to minimize the perceptuallyweighted error d_[n].e. The weighting is implemented by linear filtering.2 Perceptual Weighting of the Difference Signal Since −d[n] is the error in the reconstructed signal. Figure 7.4.14) The poles of W(z) lie at the same angle but at α2 times the radii of the poles of H(z). i.9 and α2 = 0..14). (7. In analysisbysynthesis coding. this is accomplished by quantization noise feedback or preemphasis/ deemphasis filtering.4 are used in (7. Clearly. which is the shape desired.3. the input 7.9 shows the frequency response of such a filter. In ADPCM. 7.
which we designate here as fγ[n] for 0 ≤ n ≤ L − 1. 0 1000 2000 3000 4000 −30 −20 −10 0 10 20 30 frequency in Hz. By varying α1 and α2 in (7.14) the relative distribution of error can be adjusted. 7. such errors would be masked by the high amplitude low frequencies. Thus. i. where L is the excitation frame length and γ ranges over the finite set of components. 120 Digital Speech Coding 7. log magnitude in dB Fig..filter gain is low) and it deemphasizes the low frequencies in the error signal. The input is composed of a finite sum of scaled components selected from the given collection.3. in this case.e.3 Generating the Excitation Signal Most analysisbysynthesis systems generate the excitation from a finite fixed collection of input components. . where.4.9 Comparison of frequency responses of vocal tract filter and perceptual weighting filter in analysisbysynthesis coding. it follows that the error will be distributed in frequency so that relatively more error occurs at low frequencies.
the βs and γs are all that is needed to represent the input ˆ d[n]. (7.15) The βks and the sequences fγk [n] are chosen to minimize E= L_−1 n=0 ((x[n] − h[n] ∗ ˆ d[n]) ∗ w[n])2. the output in the current frame due to the excitation determined for previous frames is computed and denoted ˆx0[n] for 0 ≤ n ≤ L − 1.16). It is very difficult to solve simultaneously for the optimum βs and γs that minimize (7. Since the component sequences are assumed to be known at both the coder and decoder. satisfactory results are obtained by solving for the component signals one at a time.16) where h[n] is the vocal tract impulse response and w[n] is the impulse response of the perceptual weighting filter with system function (7. (7. However. Normally this would be a decaying signal that could be truncated after L samples. and the perceptually weighted error would be d_ . The error signal in the current frame at this initial stage of the iterative process would be d0[n] = x[n] − ˆx0[n].14). Starting with the assumption that the excitation signal is zero during the current excitation15 analysis frame (indexed 0 ≤ n ≤ L − 1).ˆ d[n] = _N k=1 βkfγk [n].
7. assume that we have determined which of the collection of input components fγ1 [n] will reduce the weighted meansquared error the most and we have also determined the required gain β1. or if we define the perceptually weighted vocal tract impulse response as h_[n] = h[n] ∗ w[n] 15 Several analysis frames are usually included in one linear predictive analysis frame. By superposition.3 ClosedLoop Coders 121 and invoke the distributive property of convolution. subtracting it from the previously computed residual error. the corresponding equation takes the form: d _ k[n] = d _ k−1[n] − βky _ k[n].0[n] = d0[n] ∗ w[n]. The . ˆx1[n] = ˆx0[n] + β1fγ1 [n] ∗ h[n]. Generalizing to the kth iteration. This process is continued by finding the next component. and so forth. Now at the first iteration stage. then the weighted error sequence is d_ 1[n] = d_ 0[n] − β1fγ1 [n] ∗ h_[n].17) where y_[n] = fγk [n] ∗ h_[n] is the output of the perceptually weighted vocal tract impulse response due to the input component fγk [n]. The weighted error at the first iteration is d_ 1[n] = (d0[n] − β1fγ1 [n] ∗ h[n] ∗ w[n]). (7.
(7.18) is βk = L_−1 n=0 d _ k−1[n]y _ k[n] L_−1 n=0 (y _[n])2 .meansquared error at stage k of the iteration is defined as Ek = L_−1 n=0 (d _ k[n])2 = L_−1 n=0 (d _ k−1[n] − βky _ k[n])2. Assuming that γk is known. the value of βk that minimizes Ek in (7.19) and the corresponding minimum meansquared error is E (min) .18) where it is assumed that d_ k−1[n] is the weighted difference signal that remains after k − 1 steps of the process. (7.
(7.20) suggests that the meansquared error can be minimized by maximizing β2 k L_−1 n=0 (y _ k[n])2 = _ L_−1 n=0 d _ k−1[n]y _ k[n] 2 L_−1 n=0 . Equation (7.20) While we have assumed in the above discussion that fγk [n] and βk are known.k= L_−1 n=0 (d _ k−1[n])2 − β2 k L_−1 n=0 (y _ k[n])2. we have not discussed how they can be found.
3 is quite general since the only constraint was that the collection of component input sequences be finite.(y _ k[n])2 .21) and thus reduce the meansquared error the most. the component input sequences are simply isolated unit impulse sequences.4 MultiPulse Excitation Linear Prediction (MPLP) The procedure described in Section 7.3.e.21) 122 Digital Speech Coding which is essentially the normalized crosscorrelation between the new input component and the weighted residual error d_ k−1[n].4. 16 If desired. the set of input sequences so determined can be assumed and a new set of βs can be found jointly by solving a set of N linear equations. The first analysisbysynthesis coder was called multipulse linear predictive coding [4]. the complete set of component input sequences and corresponding coefficients will have been determined. i.19). In this system. where γ . and the new error sequence computed from (7..17). Once this component is found. fγ[n] = δ[n − γ]. the corresponding βk can be found from (7. An exhaustive search through the collection of input components will determine which component fγk [n] will maximize the quantity in (7.4. (7. 7.3. After N iterations of this process.
3 ClosedLoop Coders 123 In a speech coder based on multipulse analysisbysynthesis. the speech signal is represented by quantized versions of the prediction coefficients.22) where N is usually on the order of 4–5 impulses in a 5 ms (40 samples for fs = 8 kHz) excitation analysis frame.3 is therefore of the form: ˆ d[n] = _N k=1 βkδ[n − γk]. A number of different approaches have been devised for coding . 7. This is because the components y_ k[n] are all just shifted versions of h_[n]. The error reduction process can be stopped when E (min) k falls below a prescribed threshold. the pulse locations.is an integer such that 0 ≤ γ ≤ L − 1. an exhaustive search at each stage can be achieved by computing just one crosscorrelation function and locating the impulse at the maximum of the crosscorrelation. 17 It is advantageous to include two or more excitation analysis frames in one linear predictive analysis frame. 16 Alternatively. The excitation sequence derived by the process described in Section 7. N need not be fixed in advance. and the pulse amplitudes. (7.4. which normally is of 20 ms (160 samples) duration. The predictor coefficients can be coded as discussed for ADPCM by encoding any of the alternative representations either by scalar or vector quantization.3.17 In this case.
Using a preset spacing of 4 and an excitation analysis window of 40 samples yields high quality at about 10 kbps [67]. One version of RPE is the basis for the 13 kbps digital coder in the GSM mobile communications system.4. 7. it is not surprising that a great deal of effort has gone into schemes for finding excitation signals that are easier to encode than multipulse excitation. Regularpulse excitation (RPE) is a special case of multipulse excitation where it is easier to encode the impulse locations. Specifically after determining the location γ1 of the first impulse.5 CodeExcited Linear Prediction (CELP) Since the major portion of the bit rate of an analysisbysynthesis coder lies in the excitation signal. Thus. quality degrades rapidly due to the many parameters that must be coded [13]. the excitation . all other impulses are located at integer multiples of a fixed spacing on either side of γ1. however. the integer multiples. below about 10 kbps. RPE requires significantly less computation than fullblown multipulse excitation analysis. These usually involve some sort of differential coding scheme. but yet maintain high quality of reproduction of the speech signal. Because the impulse locations are fixed after the first iteration.3. only γ1. At bit rates on the order of 10–16 kbps multipulse coding can approach toll quality.the excitation parameters. The next major innovation in the history of analysisbysynthesis systems was called codeexcited linear predictive coding or CELP [112]. and the gain constants must be encoded.18 In this scheme.
Typical codebook sizes are 256. With this method. Additional bits are still required to code the βks. then a total of M · N bits will be required to code the sequences. the decoder must also have a copy of the analysis codebook so that the excitation can be regenerated at the decoder. The basic CELP framework has been applied in the development of .signal components are Gaussian random sequences stored in a 18 Earlier work by Stewart [123] used a residual codebook (populated by the Lloyd algorithm [71]) with a low complexity trellis search. Efficient searching schemes and structured codebooks have eased the computational burden. Within the codebook. the sequence lengths are typically L = 40–60 samples (5–7.”19 If the codebook contains 2M sequences. If N sequences are used to form the excitation. and modern DSP hardware can easily implement the computations in real time. 512 or 1024. then a given sequence can be specified by an Mbit number. 124 Digital Speech Coding “codebook.5 ms at 8 kHz sampling rate). This is because each codebook sequence must be filtered with the perceptually weighted impulse response before computing the cross correlation with the residual error sequence. The main disadvantage of CELP is the high computational cost of exhaustively searching the codebook at each stage of the error minimization.
numerous speech coders that operate with bit rates in the range from 4800 to 16000 bps. 19 While it may seem counterintuitive that the excitation could be comprised of white noise sequence.6 LongDelay Predictors in AnalysisbySynthesis Coders Longdelay predictors have an interesting interpretation in analysisbysynthesis . in fact.000 bps with delays less than 2 ms. The GSM half rate coder operating at 5600 bps is based on the IS54 VSELP coder.4 OpenLoop Coders 125 7.729 conjugatestructure algebraic CELP (CSACELP) coder that is used in some mobile communication systems. The ITU G. Due to the block processing structure. remember that.4. Still another standardized CELP system is the ITU G. 7. The Federal Standard 1016 (FS1016) was adopted by the Department of Defense for use in secure voice transmission at 4800 bps. Another system called VSELP (vector sum excited linear prediction). was adopted in 1989 for North American Cellular Systems. the object of linear prediction is to produce just such a signal.3. which uses multiple codebooks to achieve a bit rate of 8000 bps. CELP coders can introduce more than 40 ms delay into communication systems.728 lowdelay CELP coder uses very short excitation analysis frames and backward linear prediction to achieve high quality at a bit rate of 16.
4 OpenLoop Coders ADPCM coding has not been useful below about 9600 bps. The first step of the excitation computation would be to compute γ0 using (7. 7.21) and then β0 using (7. The additional bit rate for coding β0 and γ0 is often well worthwhile. This can be done recursively to save computation. For each value of γ0 to be tested. multipulse coding has similar limitations.coding. this requires that the weighted vocal tract impulse response be convolved with Lsample segments of the past excitation starting with sample γ0. and longdelay predictors are used in many of the standardized coders mentioned above. and CELP coding has not been used at bit rates below about 4800 bits/s. The long segment of the past excitation acts as a sort of codebook where the Lsample codebook sequences overlap by L − 1 samples. While these closedloop systems have .19). we can add a term β0 ˆ d[n − γ0] to (7. This incorporation of components of the past history of the excitation has been referred to as “selfexcitation” [104] or the use of an “adaptive codebook” [19]. Then the component β0 ˆ d[n − γ0] can be subtracted from the initial error to start the iteration process in either the multipulse or the CELP framework. If we maintain a memory of one or more frames of previous values of the excitation signal.15).
many attractive features.10 shows a source/system model for speech that is closely related to physical models for speech production. We shall refer to this as the twostate excitation model. 7. The slowlytimevarying linear system models the combined effects of vocal tract transmission. it has not been possible to generate excitation signals that can be coded at low bit rates and also produce high quality synthetic speech. We call these systems openloop systems because they do not determine the excitation by a feedback process. where the spacing between impulses is the pitch period. Unvoiced sounds are produced by exciting the system with white noise. and. P0.4. As we have discussed 126 Digital Speech Coding Fig. the excitation model can be very simple. For bit rates below 4800 bps.10 Twostateexcitation model for speech synthesis. and the gain parameter G controls the level of the .1 The TwoState Excitation Model Figure 7. the lowpass frequency shaping of the glottal pulse. in the previous chapters. in the case of voiced speech. and voiced sounds are produced by a periodic impulse train excitation. radiation at the lips. engineers have turned to the vocoder principles that were established decades ago. 7. The V/UV (voiced/unvoiced excitation) switch produces the alternating voiced and unvoiced segments of speech.
and the parameters of the linear system are supplied at periodic intervals (frames) then the model becomes a speech synthesizer.1.4. Gain. 7. operates directly on the time waveform by locating corresponding peaks and valleys in the waveform.1 Pitch. The STACF is also useful for this purpose as illustrated in Figure 4.filter output. and measuring the times between the peaks [43]. an openloop analysis/synthesis speech coder. short segments of speech are analyzed to detect periodicity (signaling voiced speech) or aperiodicity (signaling unvoiced speech). the combination of estimator and synthesizer becomes a vocoder or. P0. as we prefer. One of the simplest. When the parameters of the model are estimated directly from a speech signal. It is common to estimate the fundamental frequency F0 or equivalently. as discussed in Chapter 8.8. with time.4 OpenLoop Coders 127 the pitch period P0 = 1/F0 at a frame rate of about 50–100 times/s. which shows autocorrelation functions for both voiced . 7. To do this. The fundamental frequency varies slowly. When values for V/UV decision. more or less at the same rate as the vocal tract motions. and V/UV Detection The fundamental frequency of voiced speech can range from well below 100 Hz for lowpitched male speakers to over 250 Hz for highpitched voices of women and children. yet most effective approaches to pitch detection. G.
it is common to preprocess the speech by a spectrum flattening operation such as center clipping [96] or inverse filtering [77]. Similarly.speech. and unvoiced speech. In this case.6 shows a sequence of shorttime cepstra which moves from unvoiced to voiced speech in going from the bottom to the top of the figure. and the location of the peak is the pitch period. It should be chosen so that the shorttime energy of the synthetic output matches the shorttime energy of the input speech signal. lack of a peak in the expected range signals unvoiced speech [83]. a strong peak in the expected pitch period range signals voiced speech. which shows no evidence of periodicity. Figure 5. The time variation of the pitch period is evident in the upper plots.5. For this purpose. evincing a peak at the pitch period. the autocorrelation function value φˆn[0] at lag 0 or the cepstrum value cˆn[0] at quefrency 0 can be used to determine the . Another approach to pitch detection is suggested by Figure 5. In pitch detection applications of the shorttime autocorrelation. The gain parameter G is also found by analysis of short segments of speech. This preprocessing tends to enhance the peak at the pitch period for voiced speech while suppressing the local correlation due to formant resonances. which shows the cepstra of segments of voiced and unvoiced speech.
energy of the segment of the input signal. which is well below the bit rate used to encode the excitation signal in closedloop coders such as ADPCM or CELP.1. is simply convolved . In the original homomorphic vocoder. the impulse response was digitally coded by quantizing each cepstrum value individually (scalar quantization) [86].4. and gain must be quantized. The primary methods that have been used have been homomorphic filtering and linear predictive analysis as discussed in Chapters 5 and 6. reconstructed from the quantized cepstrum at the synthesizer. V/UV.10 can take many forms. For a frame rate of 50 frames/s. this totals 600 bps. however. For digital coding applications. Typical values are 7 bits for pitch period (P0 = 0 signals UV) and 5 bits for G. This comes at a large cost in quality of the synthetic speech output. The impulse response.2 Vocal Tract System Estimation The vocal tract system in the synthesizer of Figure 7. the pitch period. respectively. 7. Since the vocal tract filter can be coded as in ADPCM or CELP. Thus. much lower total bit 128 Digital Speech Coding rates are common in openloop systems. one cepstrum computation can yield both an estimate of pitch and the vocal tract impulse response. The Homomorphic Vocoder Homomorphic filtering can be used to extract a sequence of impulse responses from the sequence of cepstra that result from shorttime cepstrum analysis.
homomorphic filtering was used to remove excitation effects in the shorttime spectrum and then three formant frequencies were estimated from the smoothed spectra. These formant frequencies were used to control the resonance frequencies of a synthesizer comprised of a cascade of secondorder section IIR digital filters [111]. and the resulting output can overlap into the next frame. the prediction coefficients can be coded in one of the many ways that we have already discussed. the cepstrum values were coded using vector quantization. the impulse response can be changed at the time a new pitch impulse occurs. For example. i.4 OpenLoop Coders 129 excitation [6]. voicing. and gain information.with the excitation created from the quantized pitch. Such analysis/ .. In still another approach to digital coding.6 shows an example of the formants estimated for voiced speech. Figure 5.3. s[n] = Ge[n] ∗ h[n]. In this case.3.20 In a more recent application of homomorphic analysis in an analysisbysynthesis framework [21]. 7. LPC Vocoder Linear predictive analysis can also be used to estimate the vocal tract system for an openloop coder with twostate 20 Care must be taken at frame boundaries. Such a speech coder is called a formant vocoder. and the excitation derived by analysisbysynthesis as described in Section 7.4.e.
we presented Figure 7. 7.2 ResidualExcited Linear Predictive Coding In Section 7. The LPC10 system has a bit rate of 2400 bps using a frame size of 22. This system is also called LPC10 (or LPC10e) because a 10thorder covariance linear predictive analysis is used to estimate the vocal tract system. None of the methods discussed so far attempt to directly code the prediction error signal in an openloop manner.4. where the prediction error (or residual) is significantly smaller and less lowpass in nature. The twostate model attempts to construct the excitation signal by direct analysis and measurement of the input speech signal. Direct coding of the residual faces the same problem faced in . This figure shows an example of inverse filtering of the speech signal using a prediction error filter. Systems that attempt to code the residual signal directly are called residualexcited linear predictive (RELP) coders.5 ms with 12 bits/frame allocated to pitch. The speech signal can be reconstructed from the residual by passing it through the vocal tract system H(z) = 1/A(z). ADPCM and analysisbysynthesis systems derive the excitation to the synthesis filter by a feedback process.4 as motivation for the use of the source/system model in digital speech coding. voicing and gain and the remaining bits allocated to the vocal tract filter coded as PARCOR coefficients. A vocoder of this type was standardized by the Department of Defense as the Federal Standard FS1015.2.synthesis coders are called LPC vocoders.
which is quite similar to voiceexcited vocoders (VEV) [113. Adaptive delta modulation was used in [128].ADPCM or CELP: the sampling rate is the same as that of the input. and accurate coding could require several bits per sample. the problem of reducing the bit rate of the residual signal is attacked by reducing its bandwidth to about 800 Hz. In this system.11 shows a block diagram of a RELP coder [128]. 130]. but APCM could be used if the sampling rate is 130 Digital Speech Coding Inverse Filter A(z) Lowpass & Decimator Adaptive Quantizer Linear Predictive Analysis x[n] r[n] Adaptive Quantizer Decoder Interpolator & Spectrum Flattener Synthesis Filter . and coding the samples with adaptive quantization. Figure 7. lowering the sampling rate.
lowered to 1600 Hz. In the implementation of [128]. the sampling rate of the input was 6.8 kHz and the total bit rate was 9600 kbps with 6800 bps devoted to the residual signal. 7.4.3 Mixed Excitation Systems . The principle advantage of this system is that no hard V/UV decision must be made. The 800 Hz band is wide enough to contain several harmonics of the highest pitched voices. The reduced bandwidth residual is restored to full bandwidth prior to its use as an excitation signal by a nonlinear spectrum flattening operation. which restores higher harmonics of voiced speech.H(z) x[n] d[n] {α } ˆ[n] {α } d[n] Encoder Decoder Fig. The quality achieved at this rate was not significantly better than the LPC vocoder with a twostate excitation model. White noise is also added according to an empirically derived recipe. While this system did not become widely used. and no pitch detection is required. its basic principles can be found in subsequent openloop coders that have produced much better speech quality at bit rates around 2400 bps.11 Residualexcited linear predictive (RELP) coder and decoder. 7.
The weaknesses of the twostate model 7.21 In some situations. This configuration was developed as the result of careful experimentation. they each pass through a multiband spectral shaping filter. This mixed excitation helps to model shorttime spectral effects such as “devoicing” of certain bands during voiced speech. which focused onebyone on the sources of distortions manifest in the twostate excitation coder such as buzziness and tonal distortions. [75] and greatly refined by McCree and Barnwell [80].4 OpenLoop Coders 131 for excitation spurred interest in a mixed excitation model where a hard decision between V and UV is not required. The output of such systems is often described as “buzzy. errors in estimating pitch period or voicing decision cause the speech to sound unnatural if not unintelligible. the quality of the synthetic speech output leaves much to be desired. Prior to their addition.12 depicts the essential features of the mixedexcitation linear predictive coder (MELP) proposed by McCree and Barnwell [80]. The main feature is that impulse train excitation and noise excitation are added instead of switched. The gains in each of five bands are coordinated between the two filters so that the spectrum of e[n] is flat. Such a model was first proposed by Makhoul et al. a “jitter” parameter ΔP is invoked .” and in many cases. Figure 7.While twostate excitation allows the bit rate to be quite low.
132 Digital Speech Coding purpose is to reduce “peakiness” due to the minimumphase nature of the linear predictive vocal tract system. 1200. its superior performance led to a new Department of Defense standard in 1996 and subsequently to MILSTD3005 and NATO STANAG 4591. we have really only focused on a few of the most important .to better model voicing transitions.6. which operates at 2400. Several new parameters of the excitation must be estimated at analysis time and coded for transmission. and 600 bps. 7. These modifications to the basic twostate excitation LPC vocoder produce marked improvements in the quality of reproduction of the speech signal. Other important features of the MELP system are lumped into the block labeled “Enhancing Filters. 21 Devoicing is evident in frame 9 in Figure 5. The MELP coder is said to produce speech quality at 2400 bps that is comparable to CELP coding at 4800 bps. In fact.” This represents adaptive spectrum enhancement filters used to enhance formant regions. but these add only slightly to either the analysis computation or the bit rate [80].5 FrequencyDomain Coders Although we have discussed a wide variety of digital speech coding systems. 22 Such filters are also used routinely in CELP coders and are often referred to as post filters.22 and a spectrally flat “pulse dispersion filter” whose Fig. 7.12 Mixedexcitation linear predictive (MELP) decoder.
13 Subband coder and decoder for speech audio. 7. Such perfect reconstruction filter bank systems are the basis for a wideranging class of speech and audio coders called subband coders [25]. When the outputs of the filter bank are quantized by some sort of quantizer. Each of these is followed by a downsampler by a factor appropriate for the bandwidth of its bandpass input. but by careful 7. If the filters are carefully designed.13 depicts the general nature of this large class of coding systems. There is much room for variation within the general frameworks that we have identified. then it is possible for the output ˆx[n] to be virtually identical to the input x[n] [24].examples that have current relevance..5 FrequencyDomain Coders 133 Fig. for the total . These are called frequencydomain coders because they are based on the principle of decomposing the speech signal into individual frequency bands. The goal with such coders is. On the far left of the diagram is a set (bank) of analysis bandpass filters. Figure 7. On the far right in the diagram is a set of “bandpass interpolators” where each interpolator is composed of an upsampler (i. and the outputs of the downsamplers are connected to the corresponding inputs of the bandpass interpolators. design of the quantizer the output can be indistinguishable from the input perceptually. the output ˆx[n] is not equal to the input. There is one more class of coders that should be discussed. of course. raising the sampling rate by the same factor as the corresponding downsampling box) followed by a bandpass filter similar to the analysis filter.e.
suppose first that the individual channels are quantized independently. To see how subband coders work.composite bit rate23 to be as low as possible while maintaining high quality. The coupling between points on the basilar membrane results in the masking effects that we mentioned in Chapter 3. Furthermore. As discussed in Chapter 3. . the basilar membrane effectively performs a frequency analysis of the sound impinging on the eardrum. Typically. This masking was incorporated in some of the coding systems discussed so far. which incorporated the speech production model into the quantization process. but only in a rudimentary manner. Because the downsampled channel signals are full band at the lower sampling rate. Just as important. the quantization error affects all frequencies in 23 The total bit rate will be the sum of the products of the sampling rates of the downsampled channel signals times the number of bits allocated to each channel. 134 Digital Speech Coding that band. but ADPCM could be used in principle. a particular band can be quantized according to its perceptual importance. The quantizers can be any quantization operator that preserves the waveform of the signal. adaptive PCM is used. In contrast to the coders that we have already discussed. but the important point is that no other bands are affected by the errors in a given band. bands with low energy can be encoded with correspondingly low absolute error. the filter bank structure incorporates important features of the speech perception model.
6 kbps were preferred by listeners to ADPCM coders operating at 22 and 19 kbps. [25] found that subband coders operating at 16 kbps and 9. Another approach is to allocate bits among the channels dynamically according to a perceptual criterion. the quantization of the channels is done jointly. and on the basis of this a threshold of audibility is determined. bits are allocated among the channels in an iterative process so that as much of the . For example. What is not shown in that figure is additional analysis processing that is done to determine how to allocate the bits among the channels. This is the basis for modern audio coding standards such as the various MPEG standards.13. as suggested by the dotted boxes. vector quantization can be used effectively [23]. Using this audibility threshold. This typically involves the computation of a finegrained spectrum analysis using the DFT.13 if. Crochiere et al. From this spectrum it is possible to determine which frequencies will mask other frequencies. respectively. if the channel bandwidths are all the same so that the output samples of the downsamplers can be treated as a vector. Much more sophisticated quantization schemes are possible in the configuration shown in Figure 7. In an early nonuniformly spaced five channel system.The simplest approach is to preassign a fixed number of bits to each channel. Such systems are based on the principles depicted in Figure 7.
Most coding schemes have many operations and parameters that can be chosen to tradeoff among the important factors. The details of perceptual audio coding are presented in the recent textbook by Spanias [120].6 Evaluation of Coders 135 for speech production and speech perception. increasing the bit rate indefinitely does not necessarily continue to improve quality. increasing the bit rate of an . These systems rely heavily on the techniques of linear prediction. 7. In general. we have discussed a wide range of options for digital coding of speech signals. and quality as important practical dimensions of the speech coding problem. increasing the bit rate will improve quality of reproduction. and cepstrum analysis and also on models 7.6 Evaluation of Coders In this chapter. For example. processing delay. Throughout this chapter we have mentioned bit rates. All this knowledge and technique is brought to bear on the problem of reducing the bit rate of the speech representation while maintaining high quality reproduction of the speech signal. however. filter banks.quantization error is inaudible as possible for the total bit budget. complexity of computation. it is important to note that for many systems.
and quality requirements. we discuss systems whose goal is to convert ordinary text messages into intelligible and natural sounding synthetic speech so as to transmit information from a machine to a human user. but lowering the bit rate causes noticeable degradation. On the other hand. but this generally would come with an increase in computation and processing delay. The two fundamental processes performed by all TTS systems are text analysis (to determine the abstract underlying linguistic description of the speech) and speech synthesis (to produce the speech sounds corresponding to the . and their general structure is illustrated in Figure 8. Such systems are often referred to as texttospeech synthesis (or TTS) systems. available transmission capacity. In the final analysis.openloop twostateexcitation LPC vocoder above about 2400 bps does not improve quality very much. In other words. 8 TexttoSpeech Synthesis Methods In this chapter. The input to the TTS system is text and the output is synthetic speech. the choice of a speech coder will depend on constraints imposed by the application such as cost of coder/decoder. Fortunately.2. we will be concerned with computer simulation of the upper part of the speech chain in Figure 1. there are many possibilities to choose from. robustness to transmission errors. improved pitch detection and source modeling can improve quality in an LPC vocoder as witnessed by the success of the MELP system.1.
what rate of speaking is most appropriate for the material being spoken. normalizing the text (so as to determine how to pronounce words like proper names or . and also must use rules to determine word etymology to decide on how best to pronounce names and foreign words and phrases. Fig.1 Text Analysis The text analysis module of Figure 8. and how much emphasis should be given to individual words and phrases within the final spoken output speech (3) semantic focus and ambiguity resolution: the text analysis process must resolve homographs (words that are spelled alike but can be pronounced differently. Figure 8. namely: (1) pronunciation of the text string: the text analysis process must decide on the set of phonemes that is to be spoken. email message versus paragraph of text from an encyclopedia article).1 Block diagram of general TTS system. The first stage of processing does some basic text processing operations. 136 8.2 shows more detail on how text analysis is performed. the degree of stress at various points in speaking.g.1 Text Analysis 137 8. The input to the analysis is plain English text.1 must determine three things from the input text string. including detecting the structure of the document containing the text (e.. the intonation of the speech. 8. and the duration of each of the sounds in the utterance (2) syntactic structure of the sentence to be spoken: the text analysis process must determine where to place pauses. depending on context).text input).
138 TexttoSpeech Synthesis Methods Fig.1.1 Text Analysis 139 where. 8.homographs with multiple pronunciations). The output of the basic text processing step is tagged text. The basic text processing benefits from an online dictionary of word pronunciations along with rules for determining word etymology. Louis” Example 2: “She worked for DEC in Maynard MA” 8. ?.2 Components of the text analysis process. . and to decide their significance with regard to the sentence and paragraph structure of the input text.2 Text Normalization Text normalization methods handle a range of text problems that occur in real applications of TTS systems. a question mark. including how to handle abbreviations and acronyms as in the following sentences: Example 1: “I live on Bourbon St.5 in. or an exclamation point. and finally performing a linguistic analysis to determine grammatical information about words and phrases within the text. For example. long” where there are two periods. the text “St.1. depending on the context. However this is not always the case as in the sentence.1 Document Structure Detection The document structure detection module seeks to determine the location of all punctuation marks in the text. 8. neither of which denote the end of the sentence..” is pronounced as street or saint. !. “This car is 72. and in Example 2 the acronym DEC can . in Example 1. in St. where the tags denote the linguistic properties of the words of the input text string. an end of sentence marker is usually a period. 8.
3 Linguistic Analysis The third step in the basic text processing block of Figure 8.1. i.e. and twenty. emotional.2 is a linguistic analysis of the input text.” and dates. the following linguistic properties: • the part of speech (POS) of the word • the sense in which each word is used in the current context • the location of phrases (or phrase groups) with a sentence (or paragraph). irate. but typically a simple.. Digital Equipment Corporation. times currency.” Other examples of text normalization include number strings like “1920” which can be pronounced as the year “nineteen twenty” or the number “one thousand. A conventional parser could be used as the basis of the linguistic analysis of the printed text. etc. the use of a pronoun to refer back to another word unit) • the word (or words) on which emphasis are to be placed. but is virtually never pronounced as the letter sequence “D E C. i. where a pause in speaking might be appropriate • the presence of anaphora (e. Thus the string “$10. nine hundred.50” should be pronounced as “ten dollars and fifty cents” rather than as a sequence of characters.e.g. with the goal of determining. One other important text normalization problem concerns the pronunciation of proper names.. e. 8. for each word in the printed string.be pronounced as either the word “deck” (the spoken acronym) or the name of the company. for prominence in the sentence • the style of speaking. etc. account numbers.g.. shallow analysis is performed .. relaxed. especially those from languages other than English.
5 Homograph Disambiguation The homograph disambiguation operation must resolve the correct pronunciation of each word in the input string that has more than one pronunciation.6 LettertoSound (LTS) Conversion The second step of phonetic analysis is the process of graphemetophoneme conversion. both locally (emphasis) and globally (speaking style).1.4 Phonetic Analysis Ultimately.1.2 provides the processing that enables the TTS system to perform this conversion. 8. in the second phrase the word “absent” is a verb and the accent is on the second syllable. The basis for this is the context in which the word occurs. the tagged text obtained from the basic text processing block of a TTS system has to be converted to a sequence of tagged phones which describe both the sounds to be produced as well as the manner of speaking. One simple example of homograph disambiguation is seen in the phrases “an absent boy” versus the sentence “do you choose to absent yourself.1. namely conversion from the text to (marked) . The phonetic analysis block of Figure 8. 8. The way in which these steps are performed is as follows.since most linguistic parsers are very slow. with the help of a pronunciation dictionary. 140 TexttoSpeech Synthesis Methods 8.” In the first phrase the word “absent” is an adjective and the accent is on the first syllable.
in the word dictionary. in its entirety.1 Text Analysis 141 Fig. the dictionary search attempts to find affixes (both prefixes and suffixes) and strips them 8. . namely the sequence of speech sounds. a set of lettertosound rules is used to determine the best pronunciation (usually based on etymology of the word) of the root form of the word. 8. again followed by the reattachment of stripped out affixes (including the case of no stripped out affixes).1. their durations. If not.3 Block diagram of dictionary search for proper word pronunciation. and then does another “whole word” search. If the root form is not present in the dictionary. If so. perhaps the most straightforward method is to rely on a standard pronunciation dictionary. 8. the conversion to sounds is straightforward and the dictionary search begins on the next word. from the word attempting to find the “root form” of the word.7 Prosodic Analysis The last step in the text analysis system of Figure 8. Figure 8. First a “whole word” search is initiated to see if the printed word exists.speech sounds.3 shows the processing for a simple dictionary search for word pronunciation. Each individual word in the text string is searched independently. as is the case most often. along with a set of lettertosound rules for words outside the dictionary. Although there are a variety of ways of performing this analysis.2 is prosodic analysis which provides the speech synthesizer with the complete set of synthesis controls.
The second generation of speech synthesis methods (from 1977 to 1992) was based primarily on an LPC representation of subword units such as diphones (half phones). . is given in Figure 8. The determination of the sequence of speech sounds is mainly performed by the phonetic analysis step as outlined above. The assignment of duration and pitch contours is done by a set of pitch 142 TexttoSpeech Synthesis Methods and duration rules. The synthesis suffered from poor intelligibility and poor naturalness. it was shown that good intelligibility synthetic speech could be reliably obtained from text input by concatenating the appropriate diphone units.and an associated pitch contour (variation of fundamental frequency with time). This figure shows that there have been 3 generations of speech synthesis systems.4 Time line of progress in speech synthesis and TTS systems.2 Evolution of Speech Synthesis Systems A summary of the progress in speech synthesis. During the first generation (between 1962 and 1977) formant synthesis of phonemes using a terminal analog synthesizer was the dominant technology using rules which related the phonetic decomposition of the sentence to formant frequency contours. along with a set of rules for assigning stress and determining where appropriate pauses should be inserted so that the local and global speaking rates appear to be natural. By carefully modeling and representing diphone units via LPC parameters. 8. Although the intelligibility improved dramatically Fig. over the period 1962– 1997. 8.4.
the remaining (major) task of TTS systems is to synthesize a speech waveform whose intelligibility is very high (to make the speech useful as a means of communication between a machine and a human). in which the method of “unit selection synthesis” was introduced and perfected.8.html.1 Early Speech Synthesis Approaches Once the abstract underlying linguistic description of the text input has been determined via the steps of Figure 8. primarily by Sagisaka at ATR Labs in Kyoto [108]. Both tasks. and then discuss unit selection methods of speech synthesis later in this chapter. and whose naturalness is as close to real speech as possible. A detailed survey of progress in texttospeech conversion up to 1987 is given in the review paper by Klatt [65]. We begin our discussion of speech synthesis approaches with a review of early systems.cs. the naturalness of the synthetic speech remained low due to the inability of single diphone units to represent all possible combinations of sound using that diphone unit. The synthesis examples that accompanied that paper are available for listening at http://www.indiana. The resulting synthetic speech from this third generation technology had good intelligibility and naturalness that approached that of humangenerated speech. namely attaining high intelligibility .2 Evolution of Speech Synthesis Systems 143 over first generation formant synthesis. The third generation of speech synthesis technology was the period from 1992 to the present.edu/rhythmsp/ASA/Contents.2. 8.2.
since it does not take into account the “coarticulation” effects of producing phonemes in continuous speech.along with reasonable naturalness. etc. the words can be stored as sampled waveforms and simply concatenated in the correct sequence. Words spoken in continuous speech sentences are generally much shorter in duration than when spoken in isolation (often up to 50% shorter) as illustrated in Figure 8. or syllables (2) choice of synthesis parameters: including LPC features. articulatory parameters. sinusoidal parameters.2 Word Concatenation Synthesis Perhaps the simplest approach to creating a speech utterance corresponding to a given text string is to literally splice together prerecorded words corresponding to the desired utterance. This figure shows wideband spectrograms for the sentence “This shirt is red” spoken as a sequence of . diphones.5. and it does not provide either for the adjustment of phoneme durations or the imposition of a desired pitch variation across the utterance. (3) method of computation: including rulebased systems or systems which rely on the concatenation of stored speech units. waveform templates. are difficult to achieve and depend critically on three issues in the processing of the speech synthesizer “backend” of Figure 8. dyads. 144 TexttoSpeech Synthesis Methods 8. but unnatural sounding speech. For greatest simplicity.1. phones. namely: (1) choice of synthesis units: including whole words. This approach generally produces intelligible.2. formants.
2 Evolution of Speech Synthesis Systems 145 example.) such as employed in the speech coders discussed in Chapter 7 [102]. can be manipulated so as to blend the words together. the duration of the continuous sentence is on the order of half that of the isolated word version. The word concatenation approach can be made more sophisticated by storing the vocabulary words in a parametric form (formants. and further the formant tracks of the continuous sentence do not look like a set of uniformly compressed formant tracks from the individual words. distinct pauses between words) as shown at the top of Figure 8. This requires that the control parameters for all the words in the task vocabulary (as obtained from a training set of words) be stored as representations of the words. A special set of word concatenation rules is then used to create the control signals for the synthesizer.isolated words (with short.5 Wideband spectrograms of a sentence spoken as a sequence of isolated words (top panel) and as a continuous speech utterance (bottom panel). as shown at the bottom of Figure 8. and impose a desired pitch variation. while they are merged in the lower plot. Although such a synthesis system .5. even for this trivial Fig. The rational for this is that the parametric representations. etc. shorten them. the boundaries between words in the upper plot are sharply defined. LPC parameters. 8.5. being more closely related to the model for speech production. It can be seen that. 8. and as a continuous utterance. Furthermore.
One approach was based on controlling the motions of a physical model of the speech articulators based on the sequence of phonemes from the text analysis. 146 TexttoSpeech Synthesis Methods rules that are mostly derived empirically.g. shorter units such as phonemes or diphones are more suitable for synthesis. For example. This requires sophisticated control 1 It should be obvious that whole word concatenation synthesis (from stored waveforms) is also impractical for general purpose synthesis of speech. Efforts to overcome the limitations of word concatenation followed two separate paths. As we will see.would appear to be an attractive alternative for general purpose synthesis of speech. but a major problem is that there are far too many words to store in a word catalog for word concatenation synthesis to be practical except in highly restricted situations.. From the vocal tract shapes and sources produced by the articulatory model. control parameters (e. in reality this type of synthesis is not a practical approach.7 million distinct surnames in the United States and each of them would have to be spoken and stored for a general word concatenation synthesis method. there are about 1.1 There are many reasons for this. A second. equally important limitation is that wordlength segments of speech are simply the wrong subunits. formants and pitch) can be derived by applying the acoustic theory .
velum.). either via direct methods (namely solving the wave equation). An alternative approach eschews the articulatory model and proceeds directly to the computation of the control signals for a source/system model (e. Such models were thought to be inherently more natural than vocal tract analog models since: • we could impose known and fairly well understood physical constraints on articular movements to create realistic motions of the tongue.2. pitch period. etc. or indirectly by converting articulatory shapes to formants or LPC parameters • we could highly constrain the motions of the articulatory parameters so that only natural motions would occur.of speech production and then used to control a synthesizer such those used for speech coding. the rules (algorithm) for computing the control parameters are mostly derived by empirical means. • we could use Xray data (MRI data today) to study the motion of the articulators in the production of individual speech sounds. thereby potentially making the speech more natural sounding. Again. LPC parameters. 8. jaw.3 Articulatory Methods of Synthesis Articulatory models of human speech are based upon the application of acoustic theory to physical models such as depicted in highly stylized form in Figure 2. thereby increasing our understanding of the dynamics of speech production • we could model smooth articulatory parameter motions between sounds. etc.1 [22]. What we have learned about articulatory modeling of speech is . teeth. formant parameters..g.
This synthesis process has been called terminal analog synthesis because it is based on a model (analog) of the human vocal tract production of speech that seeks to produce a signal at its output terminals that is equivalent to the signal produced by a human talker. and thus. articulatory speech synthesis methods have not been found to be practical for synthesizing speech of acceptable quality. Speech is produced by varying (in time) the excitation and the vocal tract model control parameters at a rate commensurate with the sounds being produced. So far we have been unable to learn all such rules. 8. namely an excitation source.2. In this approach. e[n].that it requires a highly accurate model of the vocal cords and of the vocal tract for the resulting synthetic speech quality to be considered acceptable.2 The basis for terminal analog synthesis of speech is the source/system model of speech production that we have used many times in this text.4 Terminal Analog Synthesis of Speech The alternative to articulatory synthesis is called terminal analog speech synthesis.2 Evolution of Speech Synthesis Systems 147 articulator motion in the context of the sounds being produced. and a transfer function of the human vocal tract in the form of a rational system . It further requires rules for handling the dynamics of the 8. each sound of the language (phoneme) is characterized by a source excitation function and an ideal vocal tract model.
. ap} are the (timevarying) coefficients of the vocal tract filter. E(z) is the ztransform of the vocal tract excitation signal. This can be confusing since today.b1. . i. bq} and {ak} = {a1.) 2 In the designation “terminal analog synthesis”. and {bk} = {b0.b2. the terms analog and terminal result from the historical context of early speech synthesis studies.function.1) where S(z) is the ztransform of the output speech signal. . (See Figures 2. H(z) = S(z) E(z) = B(z) A(z) = b0 + _q k=1 bkz −k 1− _p k=1 akz −k . both the excitation signal . e[n]. s[n]. “analog” implies “not digital” as well as an analogous thing. .2 and 4. . “Terminal” originally implied the “output terminals” of an electronic analog (not digital) circuit or system that was an analog of the human speech production system. (8. . .1.a2. .. 148 TexttoSpeech Synthesis Methods For most practical speech synthesis systems.e.
it has been shown that it is preferable to factor the numerator and denominator polynomials into either a series of cascade (serial) resonances. For unvoiced speech a simpler model (with one complex pole and one complex zero). AV . Finally a fixed spectral compensation network can be used to model the combined effects of glottal pulse shape and radiation characteristics from the lips and mouth. 111]. and one fixed resonance F4. is shown in Figure 8.F3). P0). based on the above discussion.F2. The vocal tract representation of (8.6 Speech synthesizer based on a cascade/serial (formant) synthesis model. a timevarying voiced signal gain.1) change periodically so as to synthesize different phonemes. 8. A complete serial terminal analog speech synthesis model.1) can be implemented as a speech synthesis system using a direct form implementation. implemented via a parallel branch is adequate. It has also been shown that an allpole model (B(z) =constant) is most appropriate for voiced (nonnasal) speech. and an allpole discretetime system that consists of a cascade of 3 timevarying resonances (the first three formants. F1.6 [95. based on two real poles in the zplane.3 Unit Selection Methods 149 The unvoiced speech (lower) branch includes a white noise . Fig. or into a parallel combination of resonances. However. 8. The voiced speech (upper) branch includes an impulse generator (controlled by a timevarying pitch period.properties (P0 and V/UV ) and the filter coefficients of (8.
6 are alleviated by a more complex model proposed by Klatt [64]. it remains a very challenging task to compute the synthesis parameters. The voiced and unvoiced components are added and processed by the fixed spectral compensation network to provide the final synthetic speech output. AN. However. Many of the short comings of the model of Figure 8. a timevarying unvoiced signal gain. The resulting quality of the synthetic speech produced using a terminal analog synthesizer of the type shown in Figure 8.6 is highly variable with explicit model shortcomings due to the following: • voiced fricatives are not handled properly since their mixed excitation is not part of the model of Figure 8.generator. even with a more sophisticated synthesis model. Nevertheless Klatt’s Klattalk system achieved adequate quality by 1983 to justify commercialization by the Digital Equipment Corporation as the DECtalk system. . is inadequate and produces buzzy sounding voiced speech • the spectral compensation model is inaccurate and does not work well for unvoiced sounds. and a resonance/ antiresonance system consisting of a timevarying pole (FP ) and a timevarying zero (FZ).6 • nasal sounds are not handled properly since nasal zeros are not included in the model • stop consonants are not handled properly since there is no precise timing and control of the complex excitation signal • use of a fixed pitch pulse shape. independent of the pitch period.
if an infinite number of segments were recorded and saved. as we discussed. is to use synthesis segments that are sections of prerecorded 150 TexttoSpeech Synthesis Methods natural speech [31.Some DECtalk systems are still in operation today as legacy systems. based on unit selection methods.3 Unit Selection Methods The key idea of a concatenative TTS system.3 now provide superior quality in current applications. Concatenative speech synthesis systems. are what is conventionally known as “data driven” approaches since their performance tends to get better the more data that is used for training the system and selecting appropriate units. the resulting synthetic speech would sound natural for virtually all possible synthesis tasks. The word concatenation method discussed in Section 8. 8. although the unit selection methods to be discussed in Section 8. Ultimately. 50. shorter segments are required to achieve better quality synthesis. and saved in the database. The basic idea is that the more segments recorded.2 is perhaps the simplest embodiment of this idea.2. In order to design and build a unit selection system based on using recorded speech segments. however. using unit selection methods. several issues have to be resolved. annotated. 108]. the better the potential quality of the resulting speech synthesis. including the following: .
a large corpus of speech must be obtained from which to extract the diphone units as waveform snippets.3 Before any synthesis can be done. Subword units include syllables (about 10. phonemes (about 45 in English. the most reasonable choice for unit selection synthesis is the set of diphone units. This requires significant effort if done manually. it is necessary to prepare an inventory of units (diphones). Words are prohibitive since there are essentially an infinite number of words in English. amplitude)? We now attempt to answer each of these questions.3. and diphones (about 1500–2500 in English).(1) What speech units should be used as the basic synthesis building blocks? (2) How the synthesis units are selected (extracted) from natural speech utterances? (3) How the units are labeled for retrieval from a large database of units? (4) What signal representation should be used to represent the units for storage and reproduction purposes? (5) What signal processing methods can be used to spectrally smooth the units (at unit junctures) and for prosody modification (pitch. Based on this criterion. These units are then .000 in English).1 Choice of Concatenation Units The units for unit selection can (in theory) be as large as words and as small as phoneme units. duration. demisyllables (about 2500 8. In any case. 8.3 Unit Selection Methods 151 in English). The ideal synthesis unit is context independent and easily concatenates with other (appropriate) subword units [15]. but they are highly context dependent).
represented in some compressed form for efficient storage. duration adjustment and pitch modification. the phrase “I want” would be converted to diphone units as follows: Text Input: I want. High quality coding such as MPLP or CELP is used to limit compression artifacts. Following the process of text analysis into phonemes and prosody. all of which take place on the time waveform. the phoneme sequence is first converted to the appropriate sequence of units from the inventory.3. 8. the diphones are decoded into waveforms for final merging. Since each diphone unit occurs many times in the stored inventory. The second step in online synthesis is to select the most appropriate sequence of diphone units from the stored inventory. Phonemes: /#/ AY/ /W/ /AA/ /N/ /T/ /#/ Diphones: /# AY/ /AYW/ /WAA/ /AAN/ /NT/ /T.#/. where the symbol # is used to represent silence (at the beginning and end of each sentence or phrase). the selection 3A diphone is simply the concatenation of two phonemes that are allowed by the language constraints to be sequentially contiguous in natural speech. For example.2 From Text to Diphones The synthesis procedure from subword units is straightforward. At the final synthesis stage. For example. a diphone based . but far from trivial [14].
and then automatically finding the sequence of units in the database that most closely match these features. The task of the unit selection module is to choose one of each of the multiple representations of the sounds. Hence there are many versions of /HH/ and many versions of /EH/ etc. with the goal being .upon the phonemes AY and W would be denoted AY−W where the − denotes the joining of the two phonemes. 152 TexttoSpeech Synthesis Methods of the best sequence of diphone units involves solving a dynamic programming search for the sequence of units that minimizes a specified cost function. where the diphone match is defined in terms of spectral matching characteristics. having been extracted from different phonemic environments.3 Unit Selection Synthesis The “Unit Selection” problem is basically one of having a given set of target features corresponding to the spoken text. 8.4 As shown..3. as illustrated in Figure 8. This problem is illustrated in Figure 8. and possibly phase matching characteristics.7 which shows a target feature set corresponding to the sequence of sounds (phonemes for this trivial example) /HH/ /EH/ /L/ /OW/ from the word /hello/. pitch matching characteristics. The cost function generally is based on diphone matches at each of the boundaries between diphones.7. each of the phonemes in this word have multiple representations (units) in the inventory of sounds.
. 8.8 Online unit selection based on a Viterbi search through a lattice of alternatives. based on spectral. 8.) 4 If diphones were used. By specifying costs associated with each unit.to minimize the total perceptual distance between segments of the chosen sequence. both globally across the unit. pitch and phase differences throughout the sounds and especially at the boundary between sounds. 8.7 Illustration of basic process of unit selection. the sequence would be /HHEH/ /EHL/ /LOW/. (After Dutoit [31]. HH EH L OW ••• HH EH L OW “Trained” Perceptual Distance HH HH HH EH EHEH EH EH L L LL L OW OW OW Fig.3 Unit Selection Methods 153 Fig.
and the unit concatenative distortion (UCD) which is defined as the spectral (and/or pitch and/or phase) discontinuity across the boundaries of the concatenated units.and locally at the unit boundaries with adjacent units. By way of example. and a source context of the phoneme /AH/ obtained from the 154 TexttoSpeech Synthesis Methods Fig. This dynamic programming search process is illustrated in Figure 8. This optimal sequence (or equivalently the optimal path through the combination of all possible versions of each unit in the string) can be found using a Viterbi search (dynamic programming) [38. namely a nodal cost based on the unit segmental distortion (USD) which is defined as the difference between the desired spectral pattern of the target (suitably defined) and that of the candidate unit. where the transitional costs (the arcs) reflect the cost of concatenation of a pair of units based on acoustic distances. 99]. Thus there are two costs associated with the Viterbi search. The .8 for a 3 unit search. consider a target context of the word “cart” with target phonemes /K/ /AH/ /R/ /T/. we can find the sequence of units that best “join each other” in the sense of minimizing the accumulated distance across the sequence of units. and the nodes represent the target costs based on the linguistic identity of the unit. The Viterbi search effectively computes the cost of every possible path through the lattice and determines the path with the lowest total cost.9 Illustration of unit selection costs associated with a string of target units and a presumptive string of selected units. 8. throughout the unit. source word “want” with source phonemes /W/ /AH/ /N/ /T/.
tj) + N_−1 j=1 dt(θj.θ2. . Figure 8. tN} is defined as: d(Θ. .θj+1). (8. The optimal path (corresponding to the optimal sequence of units) can . Θ = {θ1. . . T = {t1. tj) is the USD cost associated with matching target unit tj with selected unit θj and dt(θj. Associated with the match between units tj and θj is a USD Unit Cost and associated with the sequence of selected units. is a UCD concatenation cost.T) = _N j=1 du(θj. θj and θj+1.θj+1) is the UCD cost associated with concatenating the units θj and θj+1. it is noted that generally there is a small overlap region between concatenated units.USD distance would be the cost (specified analytically) between the sound /AH/ in the context /W/ /AH/ /N/ versus the desired sound in the context /K/ /AH/ /R/. the UCD cost goes to zero. t2. and the string of N target units.9 illustrates concatenative synthesis for a given string of target units (at the bottom of the figure) and a string of selected units from the unit inventory (shown at the top of the figure). The total cost of an arbitrary string of N selected units. It should be noted that when selected units θj and θj+1 come from the same source sentence and are adjacent units. Further. . We focus our attention on Target Unit tj and Selected Units θj and θj+1. . . as this is as natural a concatenation as can exist in the database. .2) where du(θj. θN}.
3 Unit Selection Methods 155 which the computational demand scales linearly with both the number of target and the number of concatenation units.be efficiently computed using a standard Viterbi search ([38. and the extra two features are log power and pitch. The weights. The concatenation cost between two units is essentially the spectral (and/or phase and/or pitch) discontinuity across the boundary between the units. Although there are a variety of choices .3).3) where p is the size of the spectral feature vector (typically p = 12 mfcc coefficients (melfrequency cepstral coefficients as explained in Section 5. there is no discontinuity in either spectrum or pitch in this case. Also. wk are chosen during the unit selection inventory creation phase and are optimized using a trialanderror procedure.θj+1). Clearly the definition of concatenation cost could be extended to more than a single boundary frame. as stated earlier. the concatenation cost is defined to be zero (dt(θj.θj+1) = 0) whenever units θj and θj+1 are consecutive (in the database) since. (8. by definition. The concatenation cost essentially measures a spectral plus log energy plus pitch difference between the two concatenated units at the boundary frames.θj+1) = _p+2 k=1 wkCk(θj. 99]) in 8. and is defined as: dt(θj.6. often represented as a VQ codebook vector).
for measuring the spectral/log energy/pitch discontinuity at the boundary, a common cost function is the normalized meansquared error in the feature parameters, namely: Ck(θj,θj+1) = [fθj k (m) − fθj+1 k (1)]2 σ2 k , (8.4) where fθj k (l) is the kth feature parameter of the lth frame of segment θj , m is the (normalized) duration of each segment, and σ2 k is the variance of the kth feature vector component. The USD or target costs are conceptually more difficult to understand, and, in practice, more difficult to instantiate. The USD cost associated with units θj and tj is of the form: du(θj, tj) = _q i=1 wt iφi % Ti(fθj i ),Ti(ftj i) & , (8.5) 156 TexttoSpeech Synthesis Methods where q is the number of features that specify the unit θj or tj , wt i, i = 1,2, . . . , q is a trained set of target weights, and Ti(·) can be
either a continuous function (for a set of features, fi, such as segmental pitch, power or duration), or a set of integers (in the case of categorical features, fi, such as unit identity, phonetic class, position in the syllable from which the unit was extracted, etc.). In the latter case, φi can be looked up in a table of distances. Otherwise, the local distance function can be expressed as a simple quadratic distance of the form φi % Ti(fθj i ),Ti(ftj i) & = ' Ti(fθj i ) − Ti(ftj i) (2 . (8.6) The training of the weights, wt i , is done offline. For each phoneme in each phonetic class in the training speech database (which might be the entire recorded inventory), each exemplar of each unit is treated as a target and all others are treated as candidate units. Using this training set, a leastsquares system of linear equations can be derived from which the weight vector can be solved. The details of the weight
training methods are described by Schroeter [114]. The final step in the unit selection synthesis, having chosen the optimal sequence of units to match the target sentence, is to smooth/modify the selected units at each of the boundary frames to better match the spectra, pitch and phase at each unit boundary. Various smoothing methods based on the concepts of time domain harmonic scaling (TDHS) [76] and pitch synchronous overlap add (PSOLA) [20, 31, 82] have been proposed and optimized for such smoothing/modification at the boundaries between adjacent diphone units. 8.4 TTS Applications Speech technology serves as a way of intelligently and efficiently enabling humans to interact with machines with the goal of bringing down cost of service for an existing service capability, or providing new products and services that would be prohibitively expensive without the automation provided by a viable speech processing interactive 8.5 TTS Future Needs 157 system. Examples of existing services where TTS enables significant cost reduction are as follows: • a dialog component for customer care applications • a means for delivering text messages over an audio connection (e.g., cellphone) • a means of replacing expensive recorded Interactive Voice Response system prompts (a service which is particularly valuable when the prompts change often during the course of the day, e.g., stock price quotations). Similarly some examples of new products and services which are enabled by a viable TTS technology are the following:
• locationbased services, e.g., alerting you to locations of restaurants, gas stations, stores in the vicinity of your current location • providing information in cars (e.g., driving directions, traffic reports) • unified messaging (e.g., reading email and fax messages) • voice portal providing voice access to webbased services • ecommerce agents • customized News, Stock Reports, Sports scores, etc. • giving voice to small and embedded devices for reporting information and alerts. 8.5 TTS Future Needs Modern TTS systems have the capability of producing highly intelligible, and surprisingly natural speech utterances, so long as the utterance is not too long, or too syntactically complicated, or too technical. The biggest problem with most TTS systems is that they have no idea as to how things should be said, but instead rely on the text analysis for emphasis, prosody, phrasing and all socalled suprasegmental features of the spoken utterance. The more TTS systems learn how to produce contextsensitive pronunciations of words (and phrases), the more natural sounding these systems will become. Hence, by way of example, 158 TexttoSpeech Synthesis Methods the utterance “I gave the book to John” has at least three different semantic interpretations, each with different emphasis on words in the utterance, i.e., I gave the book to John, i.e., not to Mary or Bob. I gave the book to John, i.e., not the photos or the
apple. I gave the book to John, i.e., I did it, not someone else. The second future need of TTS systems is improvements in the unit selection process so as to better capture the target cost for mismatch between predicted unit specification (i.e., phoneme name, duration, pitch, spectral properties) and actual features of a candidate recorded unit. Also needed in the unit selection process is a better spectral distance measure that incorporates human perception measures so as to find the best sequence of units for a given utterance. Finally, better signal processing would enable improved compression of the units database, thereby making the footprint of TTS systems small enough to be usable in handheld and mobile devices. 9 Automatic Speech Recognition (ASR) In this chapter, we examine the process of speech recognition by machine, which is in essence the inverse of the texttospeech problem. The driving factor behind research in machine recognition of speech has been the potentially huge payoff of providing services where humans interact solely with machines, thereby eliminating the cost of live agents and significantly reducing the cost of providing services. Interestingly, as a side benefit, this process often provides users with a natural and convenient way of accessing information and services. 9.1 The Problem of Automatic Speech Recognition The goal of an ASR system is to accurately and efficiently convert
To express that thought. which has not yet been achieved.a speech signal into a text message transcription of the spoken words. the transducer or microphone). which is a simplified version of the speech chain shown in Figure 1.1. quiet office. 159 160 Automatic Speech Recognition (ASR) Fig. That is. the speaker sends appropriate control signals to the articulatory speech organs which form a speech utterance whose sounds are those required to speak the desired sentence. is to perform as well as a human listener.. Once the words are chosen.2. 9. or the acoustic environment in which the speaker is located (e.g. W. the speaker must compose a linguistically meaningful sentence. outdoors). in the form of a sequence of words (possibly with pauses and other acoustic events such as uh’s.1 Conceptual model of speech production and speech recognition processes. A simple conceptual model of the speech generation and speech recognition processes is given in Figure 9. resulting in the speech waveform s[n]. noisy room. um’s.e. er’s etc. independent of the device used to record the speech (i.We . the speaker’s accent. the ultimate goal.).. It is assumed that the speaker intends to express some thought as part of a process of conversing with another human or with a machine.
The input speech signal. s[n]. temporal) features. .1 and consist of an acoustic processor which analyzes the speech signal and converts it into a set of acoustic (spectral. This is in essence a digital simulation of the lower part of the speech chain diagram in Figure 1.2.2 Block diagram of an overall speech recognition system. the mel frequency cepstrum coefficients 9. is converted to the sequence of feature vectors.2 Building a Speech Recognition System 161 Fig. by the feature analysis block (also denoted spectral analysis). followed by a linguistic decoding process which makes a best (maximum likelihood) estimate of the words of the spoken sentence. X. . . . resulting in the recognized sentence ˆW . The feature vectors are computed on a framebyframe basis using the techniques discussed in the earlier chapters. which efficiently characterize the speech sounds. In particular.refer to the process of creating the speech waveform from the speaker’s intention as the Speaker Model since it reflects the speaker’s accent and choice of words to express a given thought or request. The processing steps of the Speech Recognizer are shown at the right side of Figure 9. 9. X = {x1. . Figure 9.xT }.x2.2 shows a more detailed block diagram of the overall speech recognition system.
2. ˆW that could have produced the input sequence of feature vectors. Each of the operations in Figure 9. an Ngram language model is used to compute a language model score for each proposed word string. The final block in the process is a confidence scoring process (also denoted as an utterance verification block). including the recognition word . in some cases. The remainder of this chapter is an attempt to give the flavor of what is involved in each part of Figure 9. The pattern recognition system uses a set of acoustic models (represented as hidden Markov models) and a word lexicon to provide the acoustic match score for each proposed string.2 Building a Speech Recognition System The steps in building and evaluating a speech recognition system are the following: (1) choose the feature set and the associated signal processing for representing the properties of the speech signal over time 162 Automatic Speech Recognition (ASR) (2) choose the recognition task. Also. which is used to provide a confidence score for each individual word in the recognized string. 9. extensive digital computation. The pattern classification block (also denoted as the decoding and search block) decodes the sequence of feature vectors into a symbolic representation that is the maximum likelihood string.2 involves many details and.are widely used to represent the shorttime spectral characteristics.
1 Recognition Feature Set There is no “standard” set of features for speech recognition. Instead. articulatory.3. 9. The most popular acoustic features have been the (LPCderived) melfrequency cepstrum coefficients and their derivatives. the task syntax or language model. The analog speech signal is sampled and quantized at rates between 8000 up to 20. Some of the important issues are summarized in this section. Each of these steps may involve many choices and can involve significant research and development effort. A block diagram of the signal processing used in most modern large vocabulary speech recognition systems is shown in Figure 9.000 samples/s.2. A first order (highpass) preemphasis network (1 − αz−1) is used to compensate for the speech spectral falloff at higher frequencies and approximates the inverse to the mouth transmission frequency response.vocabulary (the lexicon). and auditory features have been utilized in a range of speech recognition systems. the basic speech sounds to represent the vocabulary (the speech units). various combinations of acoustic. The preemphasized signal is next blocked into frames of N samples. Typical values for N and M correspond to frames of duration 15–40 ms. with frame shifts of 10 ms being most common. with adjacent frames spaced M samples apart. hence adjacent . and the task semantics (if any) (3) train the set of speech acoustic and language models (4) evaluate performance of the resulting speech recognition system.
2. 9.2. It is typical to use about 13 mfcc coefficients. 9. Typically the resulting feature vector is the set of cepstral coefficients. 9. the task could be recognition of a string of digits that forms a telephone number. or identification code.and secondorder derivatives.3 Block diagram of feature extraction process for feature vector consisting of mfcc coefficients and their first and second derivatives. and their first. and 13 secondorder derivative cepstral coefficients. A Hamming window is applied to each frame prior to spectral analysis using either standard spectral analysis or LPC methods.2 Building a Speech Recognition System 163 Fig. using a digits vocabulary. making a D = 39 size feature vector. 9. Some type of cepstral bias removal is often used prior to calculation of the first and second order cepstral derivatives. or a highly constrained sequence of digits that form a password. 13 firstorder cepstral derivative coefficients.frames overlap by 5–30 ms depending on the chosen values of N and M. Following (optionally used) simple noise removal methods.3 Recognition Training . the spectral coefficients are normalized and converted to melfrequency cepstral coefficients via standard analysis methods of the type discussed in Chapter 5 [27].2 The Recognition Task Recognition tasks vary from simple word and phrase recognition systems to large vocabulary conversational interfaces to machines. For example.
11] or Viterbi alignment method). The resulting statistical models form the basis for the pattern recognition operations at the heart of the ASR system. concepts such as Viterbi search are employed in the pattern recognition process as well as in training.3. As discussed in Section 9. namely acoustic model training and language model training. Language model training requires a sequence of text strings that reflect the syntax of spoken utterances for the task at hand. Acoustic model 164 Automatic Speech Recognition (ASR) training requires recording each of the model units (whole words. Generally such text training sets are created automatically (based on a model of grammar for the recognition task) or by using existing text sources. such as magazine and newspaper articles. so that training of the acoustic models first involves segmenting the spoken strings into recognition model units (via either a Baum– Welch [10. phonemes) in as many contexts as possible so that the statistical learning method can create accurate distributions for each of the model states. and then using the segmented utterances to simultaneously build model distributions for each state of the vocabulary unit models. or closed caption transcripts .There are two aspects to training models for speech recognition. Acoustic training relies on accurately labeled sequences of speech utterances which are segmented according to the transcription.
In this section. 9.3 The Decision Processes in ASR The heart of any automatic speech recognition system is the pattern classification and decision operations.g.of television news broadcasts. e. etc. i. Typically we measure word error rate and sentence (or task) error rate as a measure of recognizer performance. X.. given the measured feature vector. .4.. 9.3 The Decision Processes in ASR 165 9.3.1 Mathematical Formulation of the ASR Problem The problem of automatic speech recognition is represented as a statistical decision problem. training sets for language models can be created from databases. Other times.e. valid strings of telephone numbers can be created from existing telephone directories. there must be a reliable and statistically significant way of evaluating recognition system performance based on an independent test set of labeled utterances.2. A brief summary of performance evaluations across a range of ASR applications is given in Section 9. Specifically it is formulated as a Bayes maximum a posteriori probability (MAP) decision process where we seek to find the word string ˆW (in the task language) that maximizes the a posteriori probability P(WX) of that string.4 Testing and Performance Evaluation In order to improve the performance of any speech recognition system. we shall give a brief introduction to these important topics. 9.
(9.1) in the form: ˆW = argmax W P(XW)P(W) P(X) .2) Equation (9. W. The term P(W) is known as the “language model” and is generally denoted as PL(W) to emphasize the linguistic nature of this term. X..2) shows that the calculation of the a posteriori probability is decomposed into two terms. i. namely P(XW). For all future calculations we disregard the denominator term. since it is independent of the word sequence W which is being optimized. ˆW= .ˆW = argmax W P(WX). where the training set is as large as reasonably possible.e.1) Using Bayes’ rule we can rewrite (9. (9. one that defines the a priori probability of the word sequence. P(X). The probabilities associated with PA(XW) and PL(W) are estimated or learned from a set of training data that have been labeled by a knowledge source.2) is often written in the form of a 3step process. W. produced the feature vector. namely P(W). The term P(XW) is known as the “acoustic model” and is generally denoted as PA(XW) to emphasize the acoustic nature of this term. and the other that defines the likelihood that the word string. usually a human expert. The recognition decoding process of (9.
is a sequence of acoustic observations corresponding to each of T frames of the speech. X. the feature vector. Step 3 PA(XW) ) *+ . In order to be more explicit about the signal processing and computations associated with each of the three steps of (9. . (9. T times the frame . Step 2 is the computation of the probability associated with the linguistic model of the words in the utterance. Step 2 .argmax ) *+W. Step 1 PL(W) ) *+ . .xT }.4) where the speech signal duration is T frames (i. (9. and Step 3 is the computation associated 166 Automatic Speech Recognition (ASR) with the search through all valid sentences in the task language for the maximum likelihood sentence. As discussed above. .e. and the word sequence W. X..x2. of the form: X = {x1. .3) where Step 1 is the computation of the probability associated with the acoustic model of the speech sounds in the sentence W.3) we need to more explicit about the relationship between the feature vector.
a12.a34. aij .5) that characterizes the spectral/temporal properties of the speech signal at time t and D is the number of acoustic features in each frame. 62].6) where there are assumed to be exactly M words in the decoded string. xtD) (9. which specify the probability of making a transition from state i to state j at each frame. Similarly we can express the optimally decoded word sequence. thereby defining the time sequence of the feature vectors over the duration of the word. Usually the selftransitions. xt. .. .2. . aii are large (close to 1. as: W = w1.4 shows a simple Q = 5state HMM for modeling a whole word. the HMM is also characterized by an explicit set of state transitions. . T is an acoustic feature vector of the form: xt = (xt1.w2. . Each HMM state is characterized by a mixture density Gaussian distribution that characterizes the statistical behavior of the feature vectors within the states of the model [61.a45.. and the jump transitions.shift in ms) and each frame.xt2.2 The Hidden Markov Model The most widely used method of building acoustic models (for both phonemes and words) is the use of a statistical characterization known as Hidden Markov Models (HMMs) [33. . .a23. in the model. Figure 9.0). In addition to the statistical feature densities within states.wM. . . are small (close to 0).. 98]. W. t = 1.3. 9. 69. (9.. 97.
11]. lefttoright. 9. π = {πi.. and initial state distribution. a labeled training set of sentences (transcribed into words and subword units) is used to guide an efficient training procedure known as the Baum–Welch algorithm [10. .3 The Decision Processes in ASR 167 Fig.1 ≤ i ≤ Q} with π1 set to 1 for the “lefttoright” models of the type shown in Figure 9. An initial HMM model is used to begin the training process.1 ≤ j ≤ Q}.4 Wordbased. The heart of the training procedure for reestimating HMM model parameters using the Baum– Welch procedure is shown in Figure 9. covariances and mixture gains for the distributions in each model state.9. B = {bj(xt).e.1 ≤ i.π) with state transition matrix A = {aij .B. learn the optimal model parameters) for each word (or subword) unit.4. state observation probability density. The details of the Baum–Welch procedure are beyond the scope of this chapter but can be found in several references on Speech Recognition methods [49. HMM with 5 states. In order to train the HMM (i. The initial model can be randomly chosen or selected based on a priori knowledge of the model 1 The Baum–Welch algorithm is also widely referred to as the forward–backward method. j ≤ Q}.1 This algorithm aligns each of the various words (or subword units) with the spoken inputs and then estimates the appropriate means. 99]. The Baum–Welch method is a hillclimbing algorithm and is iterated until a stable alignment of models and speech is obtained. The complete HMM characterization of a Qstate word model (or a subword unit like a phoneme model) is generally written as λ(A.5.
which concatenates the 3state HMM for the sound /IH/ with the 3state HMM for the sound /Z/. 9.6. 9. as shown in Figure 9.6 Subwordbased HMM with 3 states. parameters. Figure 9. It is a simple matter to go from the HMM for a whole word. 9. This process is iterated until no further improvement in probabilities occurs with each new iteration. giving the word model for the word “is” (pronounced as /IH Z/).4. however once the word model has been built it can be used much the same as whole word models for training and for evaluating word . as illustrated in Fig.5 The Baum–Welch training procedure based on a given training set of utterances. A word model is made by concatenating the appropriate subword HMMs. to an HMM for a subword unit (such as a phoneme) as shown in Figure 9.3 The Decision Processes in ASR 169 Fig.168 Automatic Speech Recognition (ASR) Fig. a middle state representing the heart of the sound.7. and an ending state representing the spectral characteristics at the end of the sound. In general the composition of a word (from subword units) is specified in a word lexicon or dictionary. The iteration loop is a simple updating procedure for computing the forward and backward model probabilities based on an input speech database (the training set of utterances) and then optimizing the model parameters to give an updated HMM. This simple 3state HMM is a basic subword unit model with an initial state representing the statistical characteristics at the beginning of a sound. 9.7 Wordbased HMM for the word /is/ created by concatenating 3state subword models for the subword units /ih/ and /z/.
. subject to the constraint that the total number of feature vectors. The resulting alignment procedure is illustrated in Figure 9.. and the last feature vector. aligns with the first state in the first word model..x2.. . xT . but clearly the alignment procedure works for any size model for any word. . x1. We are now ready to define the procedure for aligning a sequence of M word models.ww.xT }. .) The procedure for obtaining the best alignment between feature vectors and model states is based on either using the Baum–Welch statistical alignment procedure (in which we evaluate the probability of every alignment path and add them up to determine the probability of . An optimal alignment procedure determines the exact best matching sequence between word model states and feature vectors such that the first feature vector. We see the sequence of feature vectors along the horizontal axis and the concatenated sequence of word states along the vertical axis. (For simplicity we show each word model as a 5state HMM in Figure 9.strings for maximizing the likelihood as part of the speech recognition process. so that every state has at least a single feature vector associated with that state.8. exceeds the total number of model states. w1. X = {x1.. aligns with the last state in the Mth word model.wM with a sequence of feature vectors.8. T.
xT }w1. i.. . 132] for which 170 Automatic Speech Recognition (ASR) Fig. . We now return to the mathematical formulation of the ASR problem and examine in more detail the three steps in the decoding Equation (9.wM (assuming each word is represented as an HMM) and perform this computation for all possible word sequences. 9.. .w2.x2. is aligned with word i(t) and HMM model state j(t) via the function wi(t) . 9..3 Step 1 — Acoustic Modeling The function of the acoustic modeling step (Step 1) is to assign probabilities to the acoustic realizations of a sequence of words. The utility of the alignment procedure of Figure 9.8 is based on the ease of evaluating the probability of any alignment path using the Baum–Welch or Viterbi procedures. .x2. 9. . or a Viterbi alignment procedure [38. xt.. given the observed acoustic vectors.. ..w2.e.. .the word string).7) If we make the assumption that each frame. we determine the single best alignment path and use the probability score along that path as the probability measure for the current word string.wM).xT } came from the word sequence W = w1.. (9. . we need to compute the probability that the acoustic vector sequence X = {x1.3 The Decision Processes in ASR 171 This calculation can be expressed as: PA(XW) = PA({x1..3.3).8 Alignment of concatenated HMM word models with acoustic feature vectors based on either a Baum–Welch or Viterbi alignment procedure.
(9.9 which shows the set of T feature vectors (frames) along the horizontal axis. we calculate the local probability PA _ xtwi(t) j(t) _ given that we know the word from which frame t came.8) where we associate each frame of X with a unique word and state. The process of assigning individual speech frames to the appropriate word model in an utterance is based on an optimal alignment process between the concatenated sequence of word models and the sequence of feature vectors of the spoken input utterance being recognized. This alignment process is illustrated in Figure 9. and the set of M . in the word sequence.7) as the product PA(XW) = _T t=1 PA ! xtwi(t) j(t) " . wi(t) j(t). Further. we can express (9.j(t) and if we assume that each frame is independent of every other frame.
which we denote as bj(xt ). We have seen in the previous section that the probability density of each state. 9. We assume that each word model is further decomposed into a set of states which reflect the changing statistical properties of the feature vectors over time for the duration of the word.words (and word model states) along the vertical axis. Using a mixture of Gaussian densities to characterize the statistical distribution of the feature vectors in each state. is learned during a training phase of the recognizer.j = 1..Ujk ].. and we denote the states as Sj. We assume that each word is represented by an Nstate HMM model. each of which corresponds to one of the words in the utterance and its set of optimally matching feature vectors. j.μjk . The optimal segmentation of these feature vectors (frames) into the M words is shown by the sequence of boxes.9) where K is the number of mixture components in the density function. and for each word. 172 Automatic Speech Recognition (ASR) Fig. of the word model. we get a statebased probability density of the form: bj(xt) = _K k=1 cjkN[xt..N. . Within each state of each word there is a probability density that characterizes the statistical properties of the feature vectors in that state.9 Illustration of time alignment process between unknown utterance feature vectors and set of M concatenated word models. (9..2.
and covariance matrix. PA(xtwi(t) j(t)). We come back to these issues later in this section. which is calculated as PA ! xtwi(t) j(t) " = bi(t) j(t)(xt).12) The computation of (9. 1 ≤ j ≤ N.11) We now return to the issue of the calculation of the probability of frame xt being associated with the j(t)th state of the i(t)th word in 9. Ujk. for mixture k in state j. j. and N is a Gaussian density function with mean vector. (9. The key point is that we assign probabilities to acoustic realizations of a sequence of words by using hidden Markov models of the acoustic .cjk is the weight of the kth mixture component in state j. 1 ≤ j ≤ N (9.10) _∞ −∞ bj(xt )dxt = 1 . in the alignment between a given word and a set of feature vectors corresponding to that word. The density constraints are: _K k=1 cjk = 1. (9. for mixture k for state j. μjk. with the constraint cjk ≥ 0.3 The Decision Processes in ASR 173 the utterance.12) is incomplete since we have ignored the computation of the probability associated with the links between word states. and we have also not specified how to determine the withinword state.
it should be clear that for any reasonable size speech recognition task. it is impractical to create a separate acoustic model for every possible word in the vocabulary since each word would have to be spoken in every possible context in order to build a statistically reliable model of the density functions of (9. for each state of the model. The parameters. and the covariance matrix. Even for modest size vocabularies of about 1000 words. Although we have been discussing acoustic models for whole words. the mean vectors. Using an independent (and orthographically labeled) set of training data.feature vectors within words. the amount of training data required for word models is excessive.9) are. the mixture weights. we “train the system” and learn the parameters of the best acoustic models for each word (or more specifically for each sound that comprises each word). according to the mixture model of (9. The use of such subword acousticphonetic models poses no real difficulties in either training or when used to build up word models and hence is the most widely used representation . The alternative to word models is to build acousticphonetic models for the 40 or so phonemes in the English language and construct the model for a word by concatenating (stringing together sequentially) the models for the constituent phones in the word (as represented in a word dictionary or lexicon).9).
consistent with the recognition task [59. all valid text strings in the language and assigning appropriate probability scores to each string.4 Step 2 — The Language Model The language model assigns probabilities to sequences of words. by hand. of a word string. . PL(W). is to enable the computation of the a priori probability. Stateoftheart systems use contextdependent phone models as the basic units of recognition [99]. 106]. 60. including: (1) statistical training from text databases transcribed from taskspecific dialogs (a learning procedure) (2) rulebased learning of the formal grammar associated with the task (3) enumerating. We now describe the way in which such a language model is built. The purpose of the language model.for building word models in a speech recognition system. either from the task at hand or from a generic database with applicability to a wide range of tasks. W. Perhaps the most popular way of constructing the language model is through the use of a statistical Ngram word grammar that is estimated from a large training set of text utterances.3. based on the likelihood of that sequence of words occurring in the context of the task being performed by the speech recognition system. 174 Automatic Speech Recognition (ASR) 9. Hence the probability of the text string W =“Call home” for a telephone number identification task is zero since that string makes no sense for the specified task. There are many ways of building Language Models for specific tasks. or grammar.
wn−2. the probability that a word wn was preceded by the pair of words (wn−1.. for example.e.wn−1) .w2.Assume we have a large text training set of wordlabeled utterances.13) = M_ n=1 PL(wnwn−1. (9.e. as PL(W) = PL(w1.wn−1. the trigram of words) consisting of (wn−2. we compute this quantity as P(wnwn−1. (Such databases could include millions or even tens of millions of text sentences.. we have a text file that identifies the words in that sentence..wn−1.wn) as it occurs . according to an Ngram language model.wn−2) = C(wn−2.wn−2)).) For every sentence in the training set..3 The Decision Processes in ASR 175 where the probability of a word occurring in the sentence only depends on the previous N − 1 words and we estimate this probability by counting the relative frequencies of Ntuples of words in the training set.. we have the basis for an Ngram language model..wn−N+1).wn−1.wn) C(wn−2.wM) (9. If we make the assumption that the probability of a word in a sentence is conditioned on only the previous N − 1 words.. to estimate word “trigram” probabilities (i. Thus we assume we can write the probability of the sentence W. Thus.wn) is the frequency count of the word triplet (i. (9.14) 9.15) where C(wn−2....
in the text training set, and C(wn−2,wn−1) is the frequency count of the word doublet (i.e., bigram of words) (wn−2,wn−1) as it occurs in the text training set. 9.3.5 Step 3 — The Search Problem The third step in the Bayesian approach to automatic speech recognition is to search the space of all valid word sequences from the language model, to find the one with the maximum likelihood of having been spoken. The key problem is that the potential size of the search space can be astronomically large (for large vocabularies and high average word branching factor language models), thereby taking inordinate amounts of computing power to solve by heuristic methods. Fortunately, through the use of methods from the field of Finite State Automata Theory, Finite State Network (FSN) methods have evolved that reduce the computational burden by orders of magnitude, thereby enabling exact maximum likelihood solutions in computationally feasible times, even for very large speech recognition problems [81]. The basic concept of a finite state network transducer is illustrated in Figure 9.10 which shows a word pronunciation network for the word /data/. Each arc in the state diagram corresponds to a phoneme in the word pronunciation network, and the weight is an estimate of the probability that the arc is utilized in the pronunciation of the word in context. We see that for the word /data/ there are four total pronunciations,
176 Automatic Speech Recognition (ASR) Fig. 9.10 Word pronunciation transducer for four pronunciations of the word /data/. (After Mohri [81].) namely (along with their (estimated) pronunciation probabilities): (1) /D/ /EY/ /D/ /AX/ — probability of 0.32 (2) /D/ /EY/ /T/ /AX/ — probability of 0.08 (3) /D/ /AE/ /D/ /AX/ — probability of 0.48 (4) /D/ /AE/ /T/ /AX/ — probability of 0.12. The combined FSN of the 4 pronunciations is a lot more efficient than using 4 separate enumerations of the word since all the arcs are shared among the 4 pronunciations and the total computation for the full FSN for the word /data/ is close to 1/4 the computation of the 4 variants of the same word. We can continue the process of creating efficient FSNs for each word in the task vocabulary (the speech dictionary or lexicon), and then combine word FSNs into sentence FSNs using the appropriate language model. Further, we can carry the process down to the level of HMM phones and HMM states, making the process even more efficient. Ultimately we can compile a very large network of model states, model phones, model words, and even model phrases, into a much smaller network via the method of weighted finite state transducers (WFST), which combine the various representations of speech and language and optimize the resulting network to minimize the number of search states
(and, equivalently, thereby minimize the amount of duplicate computation). A simple example of such a WFST network optimization is given in Figure 9.11 [81]. Using the techniques of network combination (which include network composition, determinization, minimization, and weight pushing) and network optimization, the WFST uses a unified mathematical 9.4 Representative Recognition Performance 177 Fig. 9.11 Use of WFSTs to compile a set of FSNs into a single optimized network to minimize redundancy in the network. (After Mohri [81].) framework to efficiently compile a large network into a minimal representation that is readily searched using standard Viterbi decoding methods [38]. Using these methods, an unoptimized network with 1022 states (the result of the cross product of model states, model phones, model words, and model phrases) was able to be compiled down to a mathematically equivalent model with 108 states that was readily searched for the optimum word string with no loss of performance or word accuracy. 9.4 Representative Recognition Performance The block diagram of Figure 9.1 represents a wide range of possibilities for the implementation of automatic speech recognition. In Section 9.2.1, we suggest how the techniques of digital speech analysis discussed in Chapters 4–6 can be applied to extract a sequence of feature vectors from the speech signal, and in Section 9.3 we describe the most widely used statistical pattern recognition techniques that are employed for mapping the sequence of feature vectors into a sequence
of symbols or words. Many variations on this general theme have been investigated over the past 30 years or more, and as we mentioned before, testing and evaluation is a major part of speech recognition research and development. Table 9.1 gives a summary of a range of the performance of some speech recognition and natural language understanding systems 178 Automatic Speech Recognition (ASR) Table 9.1 Word error rates for a range of speech recognition systems. Corpus Type of Vocabulary Word speech size error rate (%) Connected digit strings (TI Database) Spontaneous 11 (0–9, oh) 0.3 Connected digit strings (AT&T Mall Recordings) Spontaneous 11 (0–9, oh) 2.0 Connected digit strings (AT&T HMIHY_c ) Conversational 11 (0–9, oh) 5.0 Resource management (RM) Read speech 1000 2.0 Airline travel information system (ATIS) Spontaneous 2500 2.5 North American business (NAB & WSJ) Read text 64,000 6.6 Broadcast News Narrated News 210,000 ≈15 Switchboard Telephone 45,000 ≈27 conversation
Callhome Telephone 28,000 ≈35 conversation that have been developed so far. This table covers a range of vocabulary sizes, speaking styles, and application contexts [92]. It can be seen that for a vocabulary of 11 digits, the word error rates are very low (0.3% for a very clean recording environment for the TI (Texas Instruments) connected digits database [68]), but when the digit strings are spoken in a noisy shopping mall environment the word error rate rises to 2.0% and when embedded within conversational speech (the AT&T HMIHY (How May I Help You) c system) the word error rate increases significantly to 5.0%, showing the lack of robustness of the recognition system to noise and other background disturbances [44]. Table 9.1 also shows the word error rates for a range of DARPA tasks ranging from • read speech of commands and informational requests about a naval ships database (the resource management system or RM) with a 1000 word vocabulary and a word error rate of 2.0% • spontaneous speech input for booking airlines travel (the airline travel information system, or ATIS [133]) with a 2500 word vocabulary and a word error rate of 2.5% 9.5 Challenges in ASR Technology 179 • read text from a range of business magazines and newspapers (the North American business task, or NAB) with a vocabulary of 64,000 words and a word error rate of 6.6% • narrated news broadcasts from a range of TV news providers
and even sentences so as to maintain an intelligent dialog with a customer.000 words and a word error rate of about 35%. (the broadcast news task) with a 210. we need more reliable and accurate utterance rejection methods so we can be sure that a word needs to be repeated when poorly recognized the first time. efficiency. and finally we need better methods of confidence scoring of words. many improvements will be required in both system performance and operational performance. phrases. ASR systems fall far short of human speech perception in all but the simplest. 9.like CNBC.5 Challenges in ASR Technology So far. In the operational area we need better methods of detecting when a person is speaking to a machine and isolating the spoken input from the background. and robustness in order to utilize the technology for a wide range of tasks. . and a separate task for live telephone conversations between two family members (the call home task) with a vocabulary of 28. most constrained tasks. we need to be able to handle users talking over the voice prompts (socalled bargein conditions). Before ASR systems become ubiquitous in society.000 word vocabulary and a word error rate of about 15% • recorded live telephone conversations between two unrelated individuals (the switchboard task [42]) with a vocabulary of 45. In the system area we need large improvements in accuracy. and under a wide range of operating conditions. on a wide range of processors.000 words and a word error rate of about 27%.
Digital speech processing systems have permeated society in the form of cellular speech coders. specialized help desks etc. and speech recognition and understanding systems that handle a wide range of requests about airline flights.Conclusion We have attempted to provide the reader with a broad overview of the field of digital speech processing and to give some idea as to the remarkable progress that has been achieved over the past 4–5 decades. signal processing. There remain many difficult problems yet to be solved before digital speech processing will be considered a mature science and technology. and perception provide the tools for 180 Conclusion 181 . we have been able to create remarkable speech processing systems whose performance increases at a steady pace. Our basic understanding of the human articulatory system and how the various muscular controls come together in the production of speech is rudimentary at best. A firm understanding of the fundamentals of acoustics. and our understanding of the processing of speech in the human brain is at an even lower level. linguistics. stock price quotations. In spite of our shortfalls in understanding. synthesized speech response systems.
April 1982. which greatly improved the final result. Sadaoki Furui. no. technical advice. pp. 1558–1564. Atal. we are responsible for any weaknesses or inaccuracies that remain in this text. Yariv Ephraim. no. with the end result of improving the productivity in our work and home environments. and editorial skill were crucial at every stage of the writing.” IEEE Transactions on Communications. 11. [2] B. “Predictive coding of speech at low bit rates. 65. vol. 182 References [1] J. vol. Acknowledgments We wish to thank Professor Robert Gray. Allen and L. and Tom Quatieri for their detailed and perceptive comments. S. Of course. Luciana Ferrer. COM30. As we increase our basic understanding of speech.” Proceedings of IEEE. L. “A unified theory of shorttime spectrum analysis and synthesis. We also wish to thank Abeer Alwan. S. for inviting us to prepare this text. B.building systems that work and can be used by the general public. 4. Rabiner. R. the application systems will only improve and the pervasiveness of speech processing in our daily lives will increase dramatically. Atal and S. Hanauer. pp. November 1977. 600–614. His patience. EditorinChief of Now Publisher’s Foundations and Trends in Signal Processing. “Speech analysis and synthesis by linear prediction . [3] B.
Nayebi. John Wiley and Sons. Schroeder. H. [7] B. S.of the speech wave.” Proceedings of ICASSP. pp. 247–254. 1062–1066. “An inequality and associated maximization technique in statistical . pp. and Signal Processing. 1982.” Bell System Technical Journal. “Predictive coding of speech signals and subjective error criterion. Richardson. E. 183 184 References [10] L. “Recursive windowing for generating autocorrelation analysis for LPC analysis. Remde. Atal and M. R. pp. R. [6] B. Speech. A Computer Laboratory Textbook. October 1970. K. [5] B. 49. Baum. vol. pp.” IEEE Transactions on Acoustics. S. vol.” IEEE Transactions on Acoustics. Barnwell III. 561–580. ASSP29. and C.” Journal of the Acoustical Society of America. S. 5. no. Schroeder. ASSP27. 1996. B. 50. June 1979. R. Speech Coding. Speech and Signal Processing. “A new model of LPC exitation for producing naturalsounding speech at low bit rates. October 1981. B. “Adaptive predictive coding of speech signals. pp. pp. [8] T. Atal and M. Atal and M.” Proceedings of IEEE ICASSP. 614–617. 1973–1986. 535–538. [9] T. [4] B. April 1980. Barnwell III. S. Schroeder. vol. 1971. Atal and J. vol. “Improved quantizer for adaptive predictive coding of speech signals at low bit rates.
27.estimation for probabilistic functions of Markov processes. 1972. E. and N. G. March 1984. Weiss. T. Inc. Australia. Conkie. [16] B. W. Healy. P. [11] L. “Diphone synthesis using unit selection. R. Rosenblatt. 384–387. Bennett. [13] M. 1963.. and saphe cracking. pp. J. Bogert. pp. “Efficient computation and encoding of the multipulse excitation for LPC. K. Garten. Beutnagel. H. and J.” Third Speech Synthesis Workshop.” Bell System Technical Journal. Kabal. A.). New York: John Wiley and Sons. Jenolan Caes. pp. [14] M. . Beutnatel and A. 164–171. Syrdal. Budapest. Berouti. ed. July 1948. 3. 446–472. “A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains. vol. Mermelstein.” Annals in Mathematical Statistics. 1970.” Proceedings of ICASSP. pseudoautocovariance. vol. “The quefrency alanysis of times series for echos: Cepstrum. “Spectra of quantized signals. Hungary. Soules. [12] W. vol. P. September 1999.” Inequalities. [15] M. Conkie. R. November 1998. crosscepstrum. “Interaction of units in a unit selection database. pp. and P. Tukey.” Proceedings of Eurospeech ’99. M. Baum. 1–8. Petri.” in Proceedings of the Symposium on Time Series Analysis. (M. and A. 41.
Cox. 2.” IEEE Journal of Selected Areas in Communications. “Diphone synthesis using an overlapadd technique for speech waveform concatenation. 735– 738. H. 452–459. E. Speech and Signal Processing. pp. “Synchornized and noiserobust audio recordings during realtime MRI scans. vol. Multirate Digital Signal Processing. Nayak. [19] J. Chung and R. V. vol. 2. Seshadri. S. no. Jayant. 120. pp. V. Narayanan.” Journal of the Acoustical Society of America. 391–409. H. vol. no. [23] R. P. 1791–1794.. [22] C. Bresch.” Proceedings of International Conference on Acoustics. “New directions in subband coding. pp. and S. Tremain. 64. vol. 2015–2018. pp. Crochiere and L. October 2006. [21] J. Welch. Introduction to Wavelets and Wavelet Transforms. February 1988. May 1989. S. vol. Shoham. E..” Proceedings of IEEE ICASSP. Nielsen. . Charpentier and M. G. J. 1986. Gopinath.” Proceedings of ICASSP. 1976. PrenticeHall Inc. and N. and T. pp. Stella. “A model of articulatory dynamics and control. [18] C. Coker. W. Quackenbush. A. Y. R. 4. pp. Burrus and R. 2. March 1992. Schafer.” Proceedings of IEEE. C. [20] F. “An expandable errorpretected 4800 bps CELP coder. 6. N. [24] R. S. Gay. 117–120. Campbell Jr. Rabiner. 1983. 1998.[17] E.. “Performance evaluation of analysisbysynthesis homomorphic vocoders. PrenticeHall Inc. K. L.
S. 122–126. L. Acoustic Theory of Speech Production. [32] G. C. B.605. [27] S. [30] H.” U. . Denes and E. Dudley. The Hague: Mouton & Co. N.” Philips Research Reports. October 1976. vol. 1952. 1960.[25] R. The speech chain. 442–466. pp. 8. Princeton: Institute for Defense Analyses. August 1980. Speech and Signal Processing. 1939. [31] T. deJager. Webber. [29] P. 55. pp. pp. Patent 2. Flanagan. “The vocoder. Netherlands: Kluwer Academic Publishers. Fant. An Introduction to TexttoSpeech Synthesis. 2nd Edition. Davis and P. S. H. July 29. Ferguson. 1993. [33] J. December 1952. Mermelstein.. “Digital coding of speech in subbands. 357–366. “Hidden Markov Analysis: An Introduction. 1970. vol. D. “Delta modulation — a new method of PCM transmission using the 1unit code. W. “Differential quantization of communication signals. 1069–1085. 1980. “Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. 28.” Bell System Technical Journal. References 185 [26] C. 17.” Bell Labs Record. B. [28] F. Freeman Company. vol. 1997. A.” IEEE Transactions on Acoustics. Dutoit. Pinson. Cutler. E. Crochiere. pp. and J. no. Walter de Gruyter.361.” Hidden Markov Models for Speech.
[39] S. “Cepstral analysis technique for automatic speaker verification. its definition. [35] J. Schafer. vol. measurement and calculation. 1972. Flanagan. L. Furui. October 1933. Sondhi. 1.[34] J.” IEEE Spectrum. SpringerVerlag. 2. Ghitza. vol. Signal Processing. and K. “Audiotry nerve representation as a basis for speech processing. Furui. Speech Analysis. no. Fletcher and W. 5. pp. L. pp. Flanagan. 2. pp. Flanagan. eds. L. Rabiner. Forney. pp. W. 3. Munson. April 1981. [37] H. D. “Synthetic voices for computers. vol. R.” in Advances in Speech Signal Processing. 61. Speech. 82– 108. 268– 278. and Signal Processing.” Bell System Technical Journal. K. pp. pp. 453– . “The Viterbi algorithm. J. vol.” IEEE Proceedings. 7. [40] S.). ASSP26. vol. vol. 52–59. October 1970. no. February 1986. Ishizaka. 54. (S. Synthesis and Perception. [41] O. March 1975. “Loudness. 254–272. 22– 45. no. 485–506.” Journal of Acoustical Society of America. Furui and M. Umeda. pp. Coker.” IEEE Transactions on Acoustics.” IEEE Transactions on Acoustics Speech. L. “Synthesis of speech from a dynamic model of the vocal cords and vocal tract. no. March 1973. [38] G. Shipley. [36] J. C. H. R. L. and N. ASSP29. “Speaker independent isolated word recognition using dynamic features of speech spectrum.
46. 17–21. 2. 186 References [44] A. pt. Rabiner.” Proceedings of ICASSP 1992. Holliman.485. “Parallel processing techniques for estimating pitch period of speech in the time domain. Gorin. Sachs. pp. 4–28. J. M. 1992. Gray. E. Greefkes and K. Gold and L. pp. 3. April 1984. Parker. 57–60. 2.” in Proceedings of First European Conference on Signal Analysis and Prediction. NY: Marcel Dekker. Prague. 1991. pp.” Foundations and Trends in Communications and Information Theory. R. [42] J. C.” Philips Technical Review. Hermansky.” Proceedings of the Interactive Voice Technology for Telecommunications Applications (IVTTA). Gray. “Code modulation with digitally controlled companding for speech transmission. R. A. M. “Auditory modeling in automatic recognition of speech. 2. 335–353. [43] B. [45] R.” Journal of Acoustical Society of America. 442–448. L. G. vol. 2006. “Vector quantization. and J. 155–239. M. [48] H. pp. 517–520. and J. B. 1997. vol. no. August 1969. pp. 1970. pp. 1996. [47] J. Wilpon. . Czech Republic. “Toeplitz and circulant matrices: A review. [46] R. no. Riemens. pp. “How may I help you?. Godfrey. McDaniel. A. “SWITCHBOARD: Telephone Speech corpus for research and development.” IEEE Signal Processing Magazine.
L. Itakura and S. [50] A. 1972. S. “A statistical method for estimation of speech spectral density and formant frequencies. and H. p. Jayant. vol. vol.” Electronics and Communications in Japan.” Bell System Technical Journal. pp. pp. “Distance measure for speech recognition based on the smoothed group delay spectrum. 2001.” Proceedings of ICASSP96. PrenticeHall Inc. [53] F. “Line spectrum representation of linear predictive coefficients of speech signals. Dallas TX. Black. [55] F. no. “Synthesis of voiced sounds from a twomass model of the vocal cords. 1257– 1260. Ishizaka and J. [51] K. Saito. pp. pp. 1. Saito. Umezaki. 1996. vol. 57. 1233–1268. Hon..” Proceedings of 6th International of Congress on Acoustics. April 1987. no. 1. 373–376.W. Itakura and S. 53A. 1968. Spoken Language Processing. Hunt and A. 1970. [54] F. C17–C20. “Analysissynthesis telephony based upon the maximum likelihood method. Atlanta. 51. 535(a). Itakura and T. s35(A). 6. [52] F. Acero. A.” in Proceedings of ICASSP87. 36–43.” Bell System . pp. [56] N.” Journal of Acoustical Society of America.[49] X. Itakura. Flanagan. pp. Huang. “Adaptive delta modulation with a onebit memory. “Unit selection in a concatenative speech synthesis system using a large speech database. vol.
Furui and M. 947–954. 1997. 1980. vol. H. Sondhi. no. [65] D. “Principles of lexical language modeling for speech recognition. Jayant and P. Wilpon. References 187 [61] B. H. and J. “Maximum likelihood estimation for mixture multivariate stochastic observations of Markov chains. Juang. Jelinek. S. pp. Rabiner. pp. 2. M. H. Juang. and S. pp. S. Jelinek. 1991. Noll. R. Jayant.” in Advances in Speech Signal Processing. September 1973. 651–699. and M. [60] F. Speech and Signal Processing.” Journal of the . S.” Bell System Technical Journal. “Review of texttospeech conversion for English. 321–342. [59] F. 1985. Digital Coding of Waveforms.” AT&T Technology Journal. PrenticeHall. pp. L. “On the use of bandpass liftering in speehc recognition. (S. H. no. July 1987. [64] D. [58] N. 7.” IEEE Transactions on Acoustics. Levinson.Technical Journal. Juang. Klatt. Sondhi. Cambridge: MIT Press. Mercer. ASSP35. pp. R. 1119–1144. “Maximum likelihood estimation for multivariate mixture observations of Markov chains. Klatt.” Journal of the Acoustical Society of America. eds. L. “Adaptive quantization with one word memory. 307–309. 32. G. Roucos. Marcel Dekker.” IEEE Transactions in Information Theory. vol. 67. vol. [63] B. Statistical Methods for Speech Recognition. 6. M. [62] B. pp. no. 64. “Software for a cascade/parallel formant synthesizer. 1986. vol. H. March 1970. pp.). [57] N. E. 1235–1249. 1984. 971–995.
Sluyter. and Signal Processing. vol. K. “The sound spectrograph. Lyon. “An introduction to the application of the theory of probabilistic functions of a Markov process to automatic speech recognition.1–42.” Proceedings of ICASSP 1984. C.” IEEE Transactions on Communications. J. September 1987. 62. detection and compression in . “A computational model of filtering. COM28. and M. 1946. 737–793. Lloyd. M. Dunn. A. Lacey. 19–49.” Bell System Technical Journal. pp. Leonard. 18. [71] S. Levinson. pp. F. 84–95. [70] Y.” IEEE Transactions on Acoustics. 4.” IEEE Transactions on Information Theory. vol. Y.” Journal of the Acoustical Society of America. no. Linde. H. pp. ASSP34. and L. Sondhi. [68] R. “Regularpulse excitation: A nove approach to effective and efficient multipulse coding of speech. 28. Koenig. Buzo. vol. 2007. Theory and Practice. CRC Press. E. Speech Enhancement. pp. “A database for speakerindependent digit recognition. [72] P. January 1980. March 1982. pp. R.4. Loizou. Gray. 42. Rabiner. 1984. 1983. E. [67] P. [69] S. Kroon.11. 82. F. vol. and R. “Least square quantization in PCM. P. 1054–1063. pp. G. vol. “An algorithm for vector quantizer design. [73] R. [66] W. Deprettere. pp.11. Speech. October 1986.Acoustical Society of America. vol. and R. 1035–1074. M. L. 129–137.
” Journal of the Acoustical Society of America. pp. no.” Proceedings of IEEE. D. “Timedomain algorithms for harmonic bandwidth reduction and timescaling of pitch signals.” IRE Transactions on Information Theory. [81] M. and A. 121–133. F. Speech and Signal Processing. 27. 64. pp.” IEEE Transactions on Speech and Audio Processing. R. pp. [76] D.” Computational . Makhoul. March 1960. V. 63. Max. 1577–1581. W. H. [77] J. Paris. “Finitestate transducers in language and speech processing. 7–12. “The SIFT algorithm for fundamental frequency estimation. McCree and T. Mohri. Markel and A. “Linear prediction: A tutorial review. Speech and Signal Processing. Viswanathan. Huggins. 4. 367– 377. 1975. Linear Prediction of Speech. pp.the cochlea. no. [78] J. no. Makhoul. vol. “A mixed source model for speech compression and synthesis. Gray. “A mixed excitation LPC vocoder model for low bit rate speech coding. July 1995. AU20. December 1972. Schwarz. [74] J.” in Proceedings of IEEE International Conference on Acoustics. D. 5. Malah. vol.” IEEE Transactions on Audio and Electroacoustics. vol. Barnwell III. 1976. P. [75] J.” IEEE Transactions on Acoustics. “Quantizing for minimum distortion. IT6. Markel. 1979. 188 References [79] J. May 1982. vol. 242–250. 561–580. New York: SpringerVerlag. France. 3. pp. vol. [80] A. pp. December 1978. V. vol. 2.
Cambridge. pp. 221–228. pp. 9. MIT. 1964. vol. [87] A. pp. 1965. 9. 1999. Massachusetts. February 1969. Oppenheim. 7. R. 41. vol. Technical Report No. V. vol. vol. 1990. “Speech spectrograms using the fast Fourier transform. Schafer. no.. [85] A. 54.” Journal of the Acoustical Society of America. AU16. R. 45. Noll. vol. “Nonlinear filtering . and T. “Superposition in a class of nonlinear systems.” Bell System Technical Journal. [90] A.Linguistics.” PhD dissertation. vol. Noll. Buck. DiscreteTime Signal Processing. 293– 309. no. “A comparative study of various schemes for speech encoding. “A speech analysissynthesis system based on homomorphic filtering. 5–6. 2. 2. June 1968. pp.. V. [84] P. 432. no. 1997. 23. [86] A. R. 2. “Cepstrum pitch determination. pp. 269–312. “Homomorphic analysis of speech. Stockham Jr.” IEEE Transactions on Audio and Electroacoustics. and J. Oppenheim. PrenticeHall Inc. Also: MIT Research Lab. 1597–1614. of Electronics. Oppenehim. W. no. Oppenheim and R. W. M. Oppenheim. “Pitch synchronous waveform processing techniques for texttospeech synthesis using diphones. [82] E. Oppenheim. W. pp. August 1970. V.” Journal of the Acoustical Society of America. [83] A. February 1967. G.” Speech Communication. November 1975. no. Schafer. Charpentier.” IEEE Spectrum. V. [89] A. [88] A. vol. 57–62. V. 293–309. V. Moulines and F. Schafer.
April 1972. vol. 24–33. and Signal Processing. 7–13. AU17. Rabiner. 1995. [95] L. August 1968. February 1977. pp. 1973. vol. R. Pallett et al. pp. R. [91] M. vol. vol. Rabiner. no. Department of Electrical Engineering.” IEEE Transactions on Audio and Electroacoustics. no. Rabiner and B. 257– 286. R. March 1969. “An introduction to hidden Markov models. [93] M. 1264–1291. “A tutorial on hidden Markov models and selected applications in speech recognition. [94] T.” MS Thesis. 1989. [92] D. Glisson. pp. Juang. Speech. 5–36.of multiplied and convolved signals. no. “The 1994 benchmark tests for the ARPA spoken language program. Discretetime speech signal processing. H. no. pp. 1. References 189 [98] L. pp.” Proceedings of 1995 ARPA Human Language Technology Workshop. 56. Com20.. vol. R.” Proceedings of IEEE. “Minimum meansquared error quantization in speech. 2002. Rabiner. [96] L. “A quasionedimensional simulation for the timevarying vocal tract. pp.” IEEE Transactions on Acoustics. F. Quatieri. S. 77.” . MIT. R. 8. 2. “On the use of autocorrelation analysis for pitch detection.” IEEE Transactions on Communications. H. 1. 225–230. Prentice Hall. Paez and T. ASSP25. “A model for synthesizing speech by rule. [97] L. D. Portnoff.” IEEE Proceedings.
February 1975. no. R. R. 1541–1558. vol. PrenticeHall Inc. Schafer. L. “Computer synthesis of speech by concatenation of formant coded words. R.” Journal of the Acoustical Society of America.. W. C. 54. 2009. 4. pp. 7. P. 2. Robinson and R. 1956. vol. and J. 1993.” British Journal of Applied Physics. 453–456. “Effect of glottal pulse shape on the quality of natural vowels. Rabiner. H. PrenticeHall Inc. Juang. 166– 181. S. Flanagan. 11.. Rosenberg. [101] L. Rabiner and R. pp. [105] A. Fundamentals of Speech Recognition. vol. “The self excited vocoder — an alternate approach to toll quality at 4800 bps. (In preparation). Theory and Application of Digital Speech Processing. R. pp. Dadson.” Bell System Technical Journal. R. [102] L.IEEE Signal Processing Magazine. Rabiner and B. 1985. E. April 1986. [99] L. 43. 822–828. no. pp. Sambur. [103] D. vol. W.” Bell System Technical Journal. 297– 315. pp. Schafer. 5. Rose and T. no. February 1971. W. “An algorithm for determining the endpoints of isolated utterances. [100] L. “Two decades of statistical language modeling: Where do we go .” Proceedings of ICASSP ’86. vol. May–June 1971. Barnwell III. R. [106] R. “A redetermination of the equalloudness contours for pure tones. 50. Rosenfeld. [104] R. Rabiner and M.
pp. E. pp. Springer. 937–940. vol. Young. 1988. 2006.” Acustica. 466. Cambridge. vol. 2. [107] M. 10.” Springer Handbook of Speech Processing.from here?. Massachusetts. [113] M. vol. 679–682. 37–53. “Echo removal by discrete generalized linear filtering. Sagisaka. Blackburn. R. 1270–1278. [114] J. Schafer.” Journal of the Acoustical Society of America. 2000. R. 16. “Homomorphic systems and cepstrum analysis of speech. “Codeexcited linear prediction (CELP): Highquality speech at very low bit rates.” Proceedings of IEEE ICASSP. Schafer and L. [108] Y.5 kc/s channel. David. 2007. 47. S. Atal. R. Speech and Signal Processing. C.” IEEE Proceedings. W. February 1970. [112] M. 88. pp. D.” PhD dissertation. W. [110] R. 8. pp.” Proceedings of the International Conference on Acoustics. C. no. Rabiner.” Journal of Phonetics. 35–43. . [109] R. “System for automatic formant analysis of voiced speech. pp.” Springer Handbook of Speech Processing and Communication. 1969. MIT. 1988. vol. Schafer. 1985. 1968. Schroeder and E. “Speech synthesis by rule using an optimal selection of nonuniform synthesis units. Schroeter. 458–465. Also: MIT Research Laboratory of Electronics. B. Technical Report No. “Basic principles of speech synthesis. “A vocoder for transmitting 10 kc/s speech over a 3. 1960. W. “Rateplace and temporalplace representations of vowels in the auditory nerve and anteroventral cochlear nucleus. pp. Schroeder and B. and E. SpringerVerlag. Sachs. [111] R. no. H.
no. Burrus. 3. Urbana. B. T. 20. R. Ingebretsen. Acoustic Phonetics. 653–709. 1949. and S. [124] T. vol. G. 53. Cannon. Volkman. vol. C. K. Stevens and J. 1. C. vol. 1998. Stockham Jr. Wiley Interscience. “A joint synchrony/meanrate model of auditory speech processing. pp. [119] F.” American Journal of Psychology. “The design of trellis waveform coders. pp.” IEEE Transactions on Speech and Audio Processing.” IEEE transactions on Communications. 15–24. Audio Signal Processing and Coding. Seneff. pp. A. vol. 1988. Soong and B. Treitel. M. Smith.. J. S. Atti. p.” Journal of Phonetics. N. Fox. May 1957. Linde.” Bell System Technical Journal. Shannon and W. “Instantaneous companding of quantized signals. Sitton. 702–710.[115] S. 55–76. Weaver. 1. [121] K. [118] B. and V. and Y. Juang. 6. [123] L. T. pp. 190 References [116] C. 27– 42. “Factoring veryhighdegree polynomials. “The relation of pitch to frequency. E. 329. “Blind deconvolution . University of Illinois Press. [122] S. W. pp. Spanias. vol. Painter. 36.H. COM30. Gray. MIT Press. The Mathematical Theory of Communication. [120] A. 16. no. Stewart. April 1982. no. M.” IEEE Signal Processing Magazine. vol. Stevens. November 2003. 2007. S. January 1993. [117] G. 1940. “Optimal quantization of LSP parameters. and R.
ASSP23. Speech and Signal Processing.” IEEE Transactions on Communications. 309–321. Makhoul. 558–561. P. Wavelets and Filter Banks. 63. 1466–1474.through digital signal processing. Magill. vol. 12. [126] Y. April 1975. pp. W. Strang and T. 2. April 1979.” IEEE Transactions on Acoustics. vol. J. 3. MA: WellesleyCambridge Press. M. and Signal Processing. “The residualexcited linear prediction vocoder with transmission rate below 9. Makhoul. and Signal Processing. [129] P. K. pp. ASSP25. 4. [130] R. “A new phase unwrapping algorithm. [127] J. “Voiceexcited LPC coders for 9. October 1987. vol.” IEEE Transactions on Acoustics. 1414–1422. Viswanathan and J. no. Tribolet. June 1975. Multirate Systems and Filter Banks. [132] A. [125] G.” vol.6 kbps speech transmission. vol.” Proceedings of IEEE. T. 35. Speech. COM23. and J. 170–177. Speech. pp. 1993. Tohkura. no. Wellesley.6 kbits/s. “Quantization properties of transmission parameters in linear predictive systems.. Nguyen. 678– 692. “A weighted cepstral distance measure for speech recognition. pp. 1996. pp. no. “Error bounds for convolutional codes and an asymptotically . Russell. Viswanathan. April 1977. vol. [128] C. PrenticeHall Inc. Un and D.” IEEE Transactions on Acoustical. Vaidyanathan. Viterbi. pp. [131] V. December 1975.
” Proceedings of DARPA Speech and Natural Language Workshop. we provide the following list of journals and books as a guide for further study.” IEEE Transactions on Information Theory. Main publication of IEEE Signal Processing Society. pp. 2nd Edition. Zwicker and H. [134] E. Papers on speech and hearing as well as other areas of acoustics. Psychoacoustics. Listing the books in chronological order of publication provides some perspective on the evolution of the field. “Evaluation of the CMU ATIS system. Speech Processing Journals • IEEE Transactions on Signal Processing. Supplemental References The specific references that comprise the Bibliography of this text are representative of the literature of the field of digital speech processing. [133] W. vol. April 1967. Fastl. Published by Elsevier. pp. In addition. • Journal of the Acoustical Society of America.optimal decoding Aalgorithm. SpringerVerlag. February 1991. General publication of the American Institute of Physics. 260–269. IT13. • IEEE Transactions on Speech and Audio. A publication of the European Association for Signal Processing (EURASIP) and of the International Speech Communication Association (ISCA). Publication of IEEE Signal Processing Society that is focused on speech and audio processing. 101–105. 1990. 191 . • Speech Communication. Ward.
IEEE Press Selected Reprint Series. W. 2001. 1999. • Advances in Speech Signal Processing. Berlin. Gray. G. Second Edition. • Linear Prediction of Speech. Berlin. • DiscreteTime Speech Signal Processing. L. 1998. S. Gold and N. • Speech and Audio Signal Processing. Schafer. W. Springer Verlag.192 Supplemental References General Speech Processing References • Speech Analysis. H. John Wiley and Sons.. Prentice Hall Inc. SpringerVerlag... 1978. S. Furui. Deller. Quatieri. Second Edition. Schafer and J. Stevens. 1976. Flanagan. O’Shaughnessy. D. 1987. New York. Jr. 2000. F. 2002. D. Jr.. Marcel Dekker Inc. Marcel Dekker Inc. J. L. • Digital Speech Processing. Proakis. Markel and A. J. 1972. K. D.). • Speech Analysis. T. R. H. Sondhi. R. J. • Acoustic Phonetics. J.. and J. • DiscreteTime Processing of Speech Signals. B. Furui and M. • Digital Processing of Speech Signals. 1991. 1979. Morgan. M. Human and Machine. MIT Press. Synthesis and Perception. L. N. WileyIEEE Press. Synthesis and Recognition. Classic Reissue. Hansen. New York. Rabiner and R. . AddisonWesley. PrenticeHall Inc. • Speech Communication. Markel (eds..
W. Kleijn and K. Prentice Hall Inc. K. 2003. Nayebi. Deng and D. John Wiley and Sons. 1996. R. Paliwal. 2003. R. Benesty. P. A. Prentice Hall Inc. Rabiner and R. Gray. M. • Speech Coding and Synthesis. Jayant and P. A Computer Laboratory Textbook.• Speech Processing. Prentice Hall Inc. Noll. T. W. John Wiley and Sons. J. Schafer. Chu. 1992. • Practical Approaches to Speech Coding. M. A Dynamic and OptimizationOriented Approach. 1987. • Speech Coding. 2008. N. E. CRC Press.. 1984. Gersho and R. L. 2000. Sondhi and Y Huang (eds. 1995. • Vector Quantization and Signal Compression. M. C.. 2009. • Digital Speech: Coding for Low Bit Rate Communication Systems. Speech Coding References • Digital Coding of Waveforms.. B. Marcel Dekker. • Theory and Application of Digital Speech Processing.). P. Kluwer Academic Publishers. Kondoz. Riek. M. Speech Synthesis . Second Edition. S. Goldberg and L. • Springer Handbook of Speech Processing and Speech Communication. Elsevier. John Wiley and Sons. Springer. L. A. Supplemental References 193 • A Practical Handbook of Speech Coders. 2004. O’Shaughnessy. • Speech Coding Algorithms. W. Barnwell and K. Papamichalis.
1998. Greenwood and J. 2004. Rabiner and B. Juang. Narayanan and A. • Progress in Speech Synthesis.).. F. 1993. Cambridge University Press. Lee. SpringerVerlag. H. • TexttoSpeech Synthesis. Coleman. 1997. Bourlard and N. 1993. John Wiley and Sons. • Statistical Methods for Speech Recognition. C. MIT Press. Kluwer Academic Publishers. 1996. K. SpringerVerlag. 1999. Cambridge University Press. Taylor. J. Allen.). T. N. A. • Connectionist Speech RecognitionA Hybrid Approach. Kluwer Academic Publishers. Olive. Childers. Kluwer Academic Publisher. • Text To Speech Synthesis: New Paradigms and Advances. • Acoustics of American English. Jelinek. W. • An Introduction to TexttoSpeech Synthesis. 1987. Sproat. 1994. L. SpringerVerlag. H. K. 1996. J. P. Prentice Hall Inc. Y. S. Klatt. Morgan. Paliwal (eds. 2008. Olive and J. H. . A. Dutoit.. Alwan (eds. • Computing Prosody. F. Hunnicutt and D. S. Speech Recognition and Natural Language Processing • Fundamentals of Speech Recognition. J. 1996. • Automatic Speech and Speaker Recognition. Prentice Hall Inc.• From Text to Speech. R. VanSanten.). Hirschberg (eds. P. Soong and K. R. • Speech Processing and Synthesis Toolboxes. J. Sagisaka. P. Higuchi. D. Campbell and N.
Vary and R. Kahrs and K. D. A. H. D. John Wiley and Sons. Theory and Practice. Ltd. P. S. Prentice Hall Inc. Prentice Hall Inc. MIT Press. Coding and Error Concealment.W. P. John Wiley and Sons. Spanias. Levinson. Kluwer Academic Publishers. • Spoken Language Processing. 2006. Schutze. E. Painter and V.. Jurafsky and J. 2005. Loizou. C. 1998. T.. Martin. 2007. Audio Processing • Applications of Digital Signal Processing to Audio and Acoustics. CRC Press. 194 Supplemental References Speech Enhancement • Digital Speech Transmission. . Atti. • Speech Enhancement. Martin. X. 1999. Manning and H. Brandenburg (eds. 2000. Enhancement.).• Foundations of Statistical Natural Language Processing. • Speech and Language Processing. A. • Audio Signal Processing and Coding. Acero and H. H. John Wiley and Sons. • Mathematical Models for Speech Technology. Huang.. 2000. Hon. C. 2007.