You are on page 1of 9

Introduction to Speech Coding

IntroductiontoSpeechCoding

Taxonomy of Speech Coders


TaxonomyofSpeechCoders

Speech Coders
Waveform Coders
Time Domain:
PCM, ADPCM

Frequency Domain:
e.g. Sub-band coder,
Adaptive transform
coder

Source Coders
Linear
Predictive
Coder

Vocoder

Speech coders in Mobile Standards


SpeechcodersinMobileStandards
Standards

Service type

Speech codec

Bit rate (kbps)

GSM

Cellular

RPE-LTP

13

CD-900

Cellular

SBC

16

USDC(IS-54)
USDC(IS
54)

Cellular

VSELP

IS-95

Cellular

CELP

1.2,2.4,4.8,9.6

IS-95(PCS)

PCS

CELP

14.4

PDC

C ll l
Cellular

VSELP

4 5 6 7 11 2
4.5,6.7,11.2

CT2

Cordless

ADPCM

32

DECT

Cordless

ADPCM

32

PHS

Cordless

ADPCM

32

DCS-1800

PCS

RPE-LTP

13

PACS

PCS

ADPCM

32

Linear Predictive Code


LinearPredictiveCode

Modelspeechproductionsystem
asanautoregressivemodel:

s ( n ) = a ( k ) s ( n k ) + e( n )
k =1

Asetofmodelparametersare
determinedforeachsegmentof
speechsignals(~30ms).
p
g
(
)
Parameters{a(k);k=1:p}isfound
bysolvingaToeplitzsystemof
equations.
q

Transferfunction
S ( z)
G
H ( z) =
=
p
E( z)
1 a(k ) z k
k =1

Toencodespeech,onemay
transmitthequantized
parameters{a(k)}andGor
equivalent parameter set Often
equivalentparameterset.Often,
linespectrumparameters(LSP)
areused.
The model order is 10 in most
Themodelorderis10inmost
speechcodingstandards.

LPC Speech Coder


LPCSpeechCoder

Buffer

Voice/
Un-voice

Pitch
Analysis

Encoder

Channel

Decoder

Excitation

Syynthesizzer

LPC
filter

CELP

Codeexcitedlinearpredictive
Code
excited linear predictive
(CELP)speechcoding.
Whitenoiseinputtothelinear
predictionmodeldoesnotgive
p
ed c o
ode does o g e
satisfactoryresults:theresidue
sequencestillcontains
informationthatisimportant
tospeechsynthesis.Henceitis
necessarytosendtheresidue
toreceivingendtoo.

Tosavespace,usevector
p ,
quantization(VQ)techniqueto
encodetheresiduesequence.
Hencethenamecode
excited.
excited
InCELP,eachcodebookisa
linearvectorcontaining0or
1.Eachcodewordlengthis60
g
samples.Successivecode
wordsareoverlapped by58
samples.Thus,alinearsearch
is performed to find the best
isperformedtofindthebest
codewordsasinputtotheLPC
model.

GSM Speech Encoder


GSMSpeechEncoder
STP

Hammingg
Window

Order = 8
Short
Term
Prediction

Segmentation
20ms
Pre-emphasis
Speech
input

LPC
Inverse
Filter

Regular
g
pulse
p
excitation
(RPE)

LTP

LAR coefficients

Long
Term
Prediction

Gain, pitch

MUX
X

Pre-processing
Pre
processing

LPF

Grid
Selection

GSM Decoding
GSMDecoding
RPE
Decodingg

LTP
Synthesis
y

STP
Synthesis
y

De-Muxx

Pitch, gain
LAR Coefficients

PostProcessingg

Implementation Issues
ImplementationIssues
Tasks:

OftenimplementedinDSP
p
chipsforembedded
LPCanalysisfiltertocalculate
thecoefficients
applications(e.g.cell
Long
Longtermpredictionforpitch
term prediction for pitch
phone)
phone).
analysisneedtofinddelayD
Theparameterquantization
andgain
partneedsbitlevel
VQsearchduringCELP
VQ search during CELP
operation.
encoding Mosttime
consuming
FIRfilteringforpre
FIR filtering for pre andpost
and post
processing

You might also like