Professional Documents
Culture Documents
(45 Hours)
0. Introduction to Digital Modulation
1. Pulse Modulation:
1.1. Pulse Code Modulation (PCM)
1.2. Differential Pulse Code Modulation (DPCM)
1.3. Delta Modulation (DM)
1.4. Pulse Width Modulation (PWM)
1.5. Pulse Position Modulation (PPM)
1.6. Pulse Amplitude Modulation (PAM)
1.2. Quantization Noise in PCM
1.3 SNR in PCM
1.4. Baseband Systems.
2. Fundamentals of Vibration:
2.1. Vibration and the Acoustic Wave equation.
2.2. Transverse and Longitudinal Vibrations.
2.3. Vibrations of Plates and Membranes .
2.4. Propagation of Acoustic Waves:
2.5. Transmission of Acoustic Waves.
2.6. Dissipation of Acoustic Energy in Fluids
2.7. Radiation and Reception of Acoustic Waves
2.8. Noise and Speech.
3. Transduction:
3.1. Electromechanical Analogues
3.2. Canonical Equations.
3.3. Transmitters
3.4. Loudspeakers.
3.5. Loudspeaker Cabinets
3.6. Receivers
3.7. Microphones.
Books
Taub and Schilling, Principles of Communication Systems, McGraw Hill Book Company, New York,
1971]
Sion Haykin, Communication Systems, 4th Edition, John Wiley & Sons, 2000
Bernard Sklar, Digital Communications: Fundamentals and Applications, Prentice Hall 2000
David R. Smith, Digital Transmission Systems, 2nd Edition, Chapman and Hall 1993
Lawrence E. Kinsler, Fundamentals of Acoustics, 4th Edition, John Wiley & Sons, 2000
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 1
Modulation: Amplitude Shift Keying (ASK)
+++++++++++++++++++++++++++++++++++++++++++++++
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 3
Frequency Shift Keying - FSK:
In a frequency shift keyed transmitter the frequency is shifted by the message.
Although there could be more than two frequencies involved in an FSK signal, in this
presentation uses a binary bit stream, and so only two frequencies are involved. The
word „keyed‟ suggests that the message is of the „on-off‟ (mark-space) variety, such as
one (historically) generated by a morse key, or more likely in the present context, a
binary.
f2 Corresponds to MARK
f1 Corresponds to SPACE
Generation of FSK
Conceptually, the transmitter could consist of two oscillators (on frequencies f 1 and f2),
with only one being connected to the output at any one time. This is shown in block
diagram form in below.
The practice is for the tones f1 and f2 to bear special inter-relationships, and to be
integer multiples of the bit rate fS. This leads to the possibility of continuous phase,
which offers advantages, especially with respect to bandwidth control..
Alternatively the frequency of a single oscillator (VCO) can be switched between two
values, thus guaranteeing continuous phase - CPFSK.
The baseband approach is easily adaptable to moving Mark and Space frequencies,
without needing to change data filters.
Different data filters are still needed when the data rate changes.
A Tb A2Tb
E1 s t dt 1 cos 4 f1t dt
Tb
2
1
0 2 0 2
A Tb A2Tb
E2 s2 t dt 1 cos 4 f 2t dt
Tb
2
0 2 0 2
Define the correlation coefficient as
1 2 Tb
s t s2 t dt cos 2 f1t cos 2 f 2t dt
Tb
E1E2
0 1 Tb 0
Useful Trigonometric Identity:
sin 2 f 2 f1 Tb sin 2 f 2 f1 Tb
2 f 2 f1 Tb 2 f 2 f1 Tb
Usually either: f2 f1 Tb k an integer (chosen by design), or relatively the sum
f 2 f1 is much much larger that the difference f2 f1 . The result of either choice
is that the sum term is usually negligible, leaving the correlation coefficient to be given
as
sin 2 f 2 f1 Tb sin 2fTb
where f f 2 f1 is the frequency
2 f 2 f1 Tb 2fTb
separation between the two signals
sin 2 f 2 f1 Tb sin 2 h
That is
2 f 2 f1 Tb 2 h
Implementation of PSK
The two PSKed signals are then added to produce one of 4 signal elements. L = 4
here.
Example Constellations
+++++++++++++++++++++++++++++++++++++++++++++++++++++++
In binary signaling, the modulator produces one of two distinct signals in response to one bit of
source data at a time.
Instead of the Q function, sometimes the complementary error function erfc is used
2
erfc x e z dz
2
where
x
1
Recalling that the definition of the Q function, Q x e y /2 dy , it can be
2
2 x
However, no channel can be used for the transmission of binary digits without first
transforming the digits to waveforms that are compatible with the channel. For baseband
channels, compatible waveforms are pulses.
A baseband signal may be from one several sources. Prior to transmission, these signals need
to be formatted to a form suitable for the transmission medium
In Figure 1.1,
The conversion from binary digits to pulse waveforms takes place in the block labelled
waveform encoder, also called a baseband modulator.
The output of the waveform encoder is typically a sequence of pulses with
characteristics that correspond to the binary digits being sent.
After transmission through the channel,
o The received waveforms are detected to produce an estimate of the
transmitted digits, and then
o The final step, (reverse) formatting recovers an estimate of the source
information.
Reconstruction of m(t)
The condition f s 2W is known as the Nyquist criterion, the sampling rate f s 2W is called the
Nyquist rate and its reciprocal called the Nyquist interval.
Ideal sampling is not practical ⇒ Need practical sampling methods.
Natural Sampling
In the figure shown, h t 1 for 0 t
and h t 0 otherwise.
The pulse train p t is also known as the gating
waveform.
Natural sampling requires only an on/off gate.
The original signal m t can still be reconstructed using a lowpass filter as long as the Nyquist
criterion is satisfied.
Flat-Top Sampling
Flat-top sampling is the most popular sampling method and involves two simple operations:
sample and hold.
PULSE MODULATION
Pulse modulation techniques are still analog modulation. For digital communications of an analog
source, quantization of sampled values is needed.
Pulse modulation is the process of transmitting signals in the form of pulses (discontinuous
signals) by using special techniques, such as
• Pulse Amplitude Modulation (PAM)
• Pulse Width Modulation (PWM)
• Pulse Position Modulation (PPM)
• Pulse Code Modulation (PCM)
Analog
Signal
Timing
Pulses
PWM
PPM
where k p is the time (or position) sensitivity parameter (in sec/volt), and the pulse p t
Ts
satisfied the condition that p t 0 when t 0 and t Ts with k p m t max
2
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 29
Demodulating the PPM Signal
In the PPM sequence the k th pulse is centred at kTs k p m kTs which is estimated as t k .
t kTs
Therefore setting tk kTs k p m kTs gives m kT s k
kp
Since all the quantities on the RHS are known, it is possible to obtain a sequence of sample
values m kT s . Then an interpolating filter (actually a low-pass filter) is used to obtain an
estimate of the analog signal m t .
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
PAM transmission does not improve the noise performance over baseband modulation, but allows
multiplexing, i.e., sharing the same transmission media by different sources.
The multiplexing advantage offered by PAM comes at the expense of a larger transmission
bandwidth
Analog Signal
which agrees with the form of the PAM signal s(t) given above.
Therefore
s t m t * h t M f fs M f kf
k
s
S f M f H f S f H f fs M f kf
k
s
The amount of equalization needed in practice is usually small. For example, for low duty cycle
T / Ts 0.1 the amount of distortion is less than 0.5 per cent, and no equalization is needed.
The transmission of PAM signals imposes rather stringent requirements on the magnitude and
phase of the channel, due to the relatively narrow pulses required.
Furthermore, the noise performance of PAM transmission can never be better than ordinary
baseband signal transmission. Accordingly, for transmission over long distances, PAM
transmission remains as a means message processing for time-division multiplexing for which
conversion to other forms of pulse modulation is subsequently performed.
Digitally
Quantized Encoded
PAM signal
Mid-Treader Mid-Riser
The Encoder
A PCM communication system is represented in Figure 1.2.
The analogue signal m(t):
is sampled, and then
these samples are subjected to the operation of quantization.
The quantized samples are applied to an encoder.
The encoder responds to each such sample by the generation of a unique and identifiable
binary pulse (or binary level) pattern.
In the examples of Figure 1.3 and 1.4 (below), the pulse pattern happens to have a numerical
significance which is the same as the order assigned to the quantized levels. However, this
feature is not essential. We could have assigned any pulse pattern to any level.
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 35
A(t)
Digital
t Clock
1 2 3 4 Serial PCM
Analog
Input Output
Sampler Encoder
7 111
110
5 101
100
3 011
Sampling Pulse 010 t
1 001
Pulses t Generator 0 000
1 2 3 4
TS 1 2 3 4 011 110 101 100
Digital Signal
Figure 1.3 A 3-Bit PCM System Showing Analogue to 3-Bit Digital Conversion
The combination of the quantizer and encoder in the dashed box of Figure 1.2 is called an
analog-to-digital converter, usually abbreviated A/D converter.
The output of the A/D converter is a digitally encoded signal, which is the signal transmitted
over the communications channel in a PCM system.
The Decoder
When the digitally encoded signal arrives at the receiver (or repeater), the first operation to
be performed is the separation of the signal from the noise which has been added during the
transmission along the channel.
Such an operation is again an operation of re-quantization; hence the first block in the receiver
in Figure 1.2 is termed a quantizer
A feature which eases the burden on this quantizer is that for each pulse interval it has only to
make the relatively simple decision of whether a pulse has or has not been received or which of
two voltage levels has occurred.
Suppose the quantized sample pulses had been transmitted instead, rather than the
binary-encoded codes for such samples. Then this quantizer would have had to have yielded, in
each pulse interval, not a simple yes or no decision, but rather a more complicated determination
about which of the many possible levels had been received.
In the examples of Figures 1.3 and 1.4, if a quantized PAM signal had been transmitted, the
receiver quantizer would have to decide which of the levels 0 to 7 was transmitted, while with a
binary PCM signal the quantizer need only distinguish between two possible levels.
The relative reliability of the yes or no decision in PCM over the multivalued decision
required for quantized PAM constitutes an important advantage for PCM.
The quantized PAM signal is now reconstituted. It is then filtered to reject any frequency
components lying outside of the baseband. The final output signal m'(t) is identical with the
input m(t) except for quantization noise and the occasional error in yes-no decision making at
the receiver due to the presence of channel noise.
Example 1.2:
An analog signal, x(t), is limited in its excursions to the range - 4 to + 4 V.
The step size between quantization levels has been set at 1V.
8 quantization levels are employed (0,1,2, …, 7);
These are located at -3.5, -2.5, . . . , +3.5 V.
Assign:
The code number 0 to the level at -3.5 V,
The code number 1 to the level at -2.5 V,
The code number 2 to the level at -1.5 V,
The code number 3 to the level at -0.5 V,
The code number 3 to the level at +0.5 V, and so on,
until the level at 3.5 V, which is assigned the code number 7.
Each code number has its representation in binary form, ranging from 000 for
code number 0 to 111 for code number 7.
The ordinate in Figure 4 is
labelled with quantization levels
and their code numbers.
Note that in the example of Figure 1.4, each sample is represented by a 3 bit codeword.
For Q levels of quantization, each sample is digitally encoded into an N-bit codeword, where
N = log2Q.
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 37
A PCM signal is obtained from the quantized PAM signal by encoding each quantized sample to a
digital codeword. In binary PCM each quantized sample is digitally encoded into an N-bit binary
codeword, where N = log2Q.
Binary digits of a PCM signal can be transmitted using many efficient modulation schemes.
There are several mappings:
Natural binary coding (NBC),
Gray mapping,
Foldover binary coding (FBC), etc.
Quantization Error
2
mM
Peak-to-Peak Range of Signal M
m t mk e
mk 2
Error, e
mk 1 e
2
mk
m
mk 1 2
m t b
m2
m1
a
2
Figure 1.5 (a) Voltage excursions for a signal m(t). The range is divided into M intervals of size . The
quantization levels are located at the centre of the interval. (b) The error voltage e(t) as a function of
the instantaneous value of the signal m(t).
Normally the probability density function p(m) of the message signal m(t) will certainly not be constant.
But when the number (Q) of quantization levels is large, so that the step size is small in
comparison with the peak-to-peak range of the message signal, it is reasonable to make the
approximation that p(m) is constant within each quantization interval.
Then in each interval, set p(m) = p(k), k =1,2,…, Q, and remove each p(k) from inside the integral,
and also make the substitution x=m- mk, k =1,2,…, Q
e2 p (1) p (2) p (3)
/2
x 2 dx
/2
p (1) p (2) p (3)
3
/12
p (1) p (2) p (3) 2 /12
But since p(k ) is the probability that the signal voltage m(t) will be in the kth quantization
interval, (k =1,2,…, Q), the sum p(1) p(2) p(3) p(Q ) 1
2
Therefore, the mean-square quantization error is e2
12
Signal-to-Quantization Noise Power Ratio
Suppose the analogue signal m(t) has a voltage range from –V to +V volts, and that each interval
is occupied with equal probability. That is, suppose that the signal is unformaly distributed.
The probability density function of signal is equal to 1/(2V), and the normalized average power
V 1 V2
of the input signal is S i m 2 (t ) V
m 2 (t ) dm
2V 3
2
Recall from above that the quantization noise power is given by NQ
12
If the number of quantization levels is Q, then Q = 2V, we have, V = Q/2
Si
Thus the signal to quantization power ratio is Q2
NQ
For the Q quantization levels, the number N of bits needed will be N = log2Q.
Si
N
Equivalently, Q= 2 , so that we have, 2 2N (for a uniformly distributed signal)
NQ
Si Si
In decibels, = 10log = 10 log10(22N) = 6N
N Q dB N Q
Si
2
2
S RMS 3Q 2
3Q 2
NQ /12 SMAX / SRMS
2
F2
S MAX
where F is the crest factor of the signal.
S RMS
V
It is sometimes called the loading factor (designated by later when discussing the
overload distortion).
In decibels,
Si Si 3Q 2
= 10log = 10log 2 20log Q 20log F 4.77
N Q dB N Q F
10log 22 N 20log F 4.77
= 6.02N – 20log10(F) + 4.77
Example 1.3
A sinusoid with maximum voltage (amplitude) V/2 is to be encoded with a PCM coder having a range of V
volts. Derive an expression for the signal-to-quantization noise (distortion) ratio and determine the
number of quantization bits required to provide an S/NQ of at least 40 dB.
Si/NQ = (3/2)22N
In decibels, (Si/NQ)dB = 6N + 1.76 dB
For 40 dB, we have 6N + 1.76 40 6N 38.24 or N 38.24/6 = 6.37
The minimum value of N is 7 bits
r
0
0
m dm
1 8d12
pb m dm 8 16d1
1 d1
1/4 1 d1
0 dm 3 1/4 dm
1
0
1
1 1
d1 mdm 1 d12 1 d1
1
3
m p b m dm r2 r1 r1 r2
r2 1
d1
3
d1 pb m dm 3 d1 dm 2 1 d1 2 1
1 1 d1 1 0
1
d1 1 m
4
4
r1 r2 1 8d12 1 d1 5
d1 2d1 r1 r2 4d12 d1 0
2 8 16d1 2 4
d1 0.4478 r1 0.1717; r2 0.7239.
Max-Lloyd quantizer
Max 1 and Lloyd 2 independently designed the quantizers minimizing the mean square error
(MSQE) and developed tables for input governed by standard distribution functions such as:
Gamma,
Laplacian,
Gaussian,
Rayleigh, and
Uniform.
Types of Quantizers
Uniform – A uniform quantizer is one with equally-spaced quantization levels (i.e. the
step-size is the constant for all the levels)
Non-Uniform – A non-uniform quantizer is one with unequally-spaced quantization levels
(i.e. the step-size is different for some or all the levels)
1
J. Max, “Quantizing for minimum distortion,” IRE Trans. Inform. Theory, Vol. IT-6, pp.16-21, Mar 1960
2
S. P. Lloyd, “Least square quantization in PCM,” IEEE Trans. Inform. Theory, Vol. IT-28, pp.129-137,1982
Quantizer Design
Given the range of input x as from a L to aU and the number of output levels as Q, design the
quantizer so that the MSQE is minimized.
MSQE E x xˆ
2
aL
aU
x xˆ 2 p( x)dx
When the number of output levels Q is very large, the quantizer d approximated as follows.
Assuming that p(x) is constant over each level, p(x) ~ p(r l),
x rl 2 p( x)dx
d l 1
dl
l 0
This is minimized by setting, the derivatives to zero: 0 and 0
rk d k
x rk 1 2 p( x)dx x rk 1 2 p( x)dx
dk d k 1
d k d k d k 1 dk
Evaluating to zero, we obtain
d k rk 1 2 p(d k ) d k rk 2 p(d k ) 0
which yields
dk rk 1 dk rk
Since d k rk 1 0 and d k rk 0 , of the two solutions, only the following is valid:
dk rk 1 dk rk
and so
rk rk 1
dk
2
Also
x rk 2 p( x)dx
d k 1
rk
rk dk
x rk p( x)dx 0
d k 1
2 dk
Hence
d k 1
rk
dk
xp ( x ) dx
d k 1
dk
p ( x ) dx
This says that the output level is the centroid of the adjacent input levels.
The solution is not in closed form. To find the input levels dk, we must find rk, vice versa.
However, by iterative techniques, both dk, and rk can be determined.
If the signal is uniformly distributed (or the number of levels Q is large), p(x) can be
taken as constant in each interval, and we have:
1
d k 1 rk d k rk
3 3
p (rk )
3 rk
Equating the derivative to zero, we have
dk 1 rk 2 dk rk 2 0
From this we obtain:
dk 1 rk dk rk
and so
d k 1 d k
rk
2
Each reconstruction level rk is midway between its two adjacent decision levels dk and dk+1.
Overload Distortion
Overload distortion results when the input signal exceeds the outermost quantizer
levels (-V, V). Notice that -V = d0 and V = dQ. For such signals, we may rewrite the mean
square quantization error as follows:
Q 1
x d 0 l x rl 2 p( x)dx
d0 d l 1
p ( x)dx
2 2
e
d
l 0
x d
Q
2
p ( x)dx
dQ
where the first and the last integrals represent the overload distortion terms.
To compute the mean square error due to overload distortion (D o), the input signal pdf must be
specified. First, let us assume the pdf to be symmetric so that the overload distortion terms
are equal, and can be combined to give
D0 2 d 2
x d Q p( x)dx
Q
If the input signal has a noise like characteristic, then a Gaussian pdf can be assumed,
described by
1 x 2 / 2 2
p ( x) e
2 2
An important example of signals that are closely described by a Gaussian pdf, by virtue of the
law of large numbers, is an FDM multichannel signal.
Speech statistics are often modelled after the Laplacian pdf, given by
2 2
D0
V 2 e
x2 / 2
2
dx VeV / 2
2 2
V/
for the Gaussian input, and
2 /
D0 2eV for the Laplacian input.
If the bit errors are assumed to be independent, then pj = Pe for each of the bits in the code,
so that
N
e 2
Pe 2
4
j 1
j 1
2 Pe 4 N 1 / 3
e 2 N e 2 Pe 4 N / 3
S
S / NQ
The overall signal-to-noise ratio is then
NQ Ne 1 Ne / NQ
where is S the signal power, NQ the quantization noise power, and Ne the thermal noise power.
S 22 N
Substituting gives
N PCM 1 4 Pe 2 2 N
for the signal-to-noise ratio for a PCM system.
In this expression, the probability Pe of bit error (bit error rate) depends on the modulation
scheme used. For example, for PSK and FSK it expressions for Pe are is below in terms of the
signal parameters.
in addition to an 40
36
- intermediate region. 38
where the signal is neither weak nor Thermal Noise Limited
36 Region
strong, in which the PCM signal-to-noise
ratio is in between the above two 34
extremes. 32
30
S
20 21 22 23 24 25 26 27 28 dB
N in
In the signal-to-noise ratio given above, the probability Pe of bit error (bit error rate) depends
on the modulation scheme used. For example, for PSK and FSK it is given respectively by
1 Eb
PSK: Pe erfc
2 N0
1 Eb
FSK: Pe erfc
0.6
2 N0
1 Eb 1 1 Si
Pe erfc
PSK :
2 N0 2 erfc
2 N N0 f M
1 1 E
erfc
2 N T f
2
S M 0
N
1 E 1 0.3 Si
Pe erfc
FSK :
2 0.6 b
N0 2 erfc
N N0 f M
1 0.3 E
erfc
N T f
2
S M N 0
(SNR)Out
S 22 N dB 10 log10 22 N
N PCM 1 4 Pe 22 N Dire
ct
or
Since 10 log10 0.6 2.218 , PSK
FSK
it follows that for weak signals, the
2.2 dB
SNR curves for PSK and FSK will be
2.3dB apart, as shown below. For strong
signals, the two expressions are the
dB
same.
(SNR)in
Problems
P.1.1 A signal to be quantized has a range normalized to 1 and a probability density function p(x) = 1 -
|x| with -1 x 1.
(a) Find the quantizer step size and levels for a uniform quantizer with eight levels.
(b) Find the eight levels for a nonuniform quantizer necessary to make the quantizer levels
equiprobable.
(c) Plot the compressor characteristic for part (b).
P.1.2 In a compact-disc (CD) digital audio system, 16-bit linear PCM is used with a sampling frequency
of 44.1 kHz for each of two stereo channels.
(a) What is the resulting data rate?
(b) What is the maximum frequency allowed on the input signal?
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 48
(c) What is the maximum S/Dq ratio in dB?
(d) If music has a loading factor = 20, find the average S/Dq in dB.
(e) If the total playing time of the CD is 70 minutes, find the total number of bits stored on the
disc. Assume that error correction coding, synchronization, and other overhead bits make up one
half of the total capacity of the disc with the remaining one-half dedicated to PCM bits.
P.1.3 The bandwidth of a TV video plus audio signal is 4.5 MHz. This signal is to be converted into linear
PCM with 1024 quantizing levels. The sampling rate is to be 20 percent above the Nyquist rate.
(a) Determine the resulting bit rate.
(b) Determine the S/Dq if the quantizer loading factor is = 6.
P.1.4 A 12-bit linear PCM coder is to be used with an analog signal in the range of 10 volts.
(a) Find the size of the quantizing step.
(b) Find the mean square error of the quantizing distortion.
(c) Find the S/Dq for a full-scale sinusoidal input.
(d) Find the S/Dq for a 1 volt sinusoidal input.
P.1.1 A signal to be quantized has a range normalized to 1 and a probability density function p(x) = 1 -
|x| with -1 x 1.
(a) Find the quantizer step size and levels for a uniform quantizer with eight levels.
Solution:
For a uniform quantizer, the levels rk and the thresholds dk are equally spaced
range 2 1 Output
Step size = =
No.of levels 8 4 r7 7
8
The first level is at
5
1 7 r6
1 1 8
2 8 8 3
The last level is at r5
8
1 7 3 1 1 3
1 1
1 r4 1
1 8 1 1 Input
4 2 4 2 4
2 8 8 d3 d4 4
d0 d1 d5 d6 d7 d8
The other levels are spaced a step size d2
r3
1
8
apart. Therefore the levels are 3
2V 2 1
r2 Q 8 4
7 5 3 1 8
, , , ,
8 8 8 8 r1
5
8
1 3 5 7
, , ,
8 8 8 8 r0
7
8
(b) Find the eight levels for a nonuniform quantizer necessary to make the quantizer levels equiprobable.
1
x dx
d j 1
Solution: We need Pj dj
pX
8
, j 0,1, 2..., 7
4 2 2 2
2 2
Taking the smaller root gives d 6 0.292893
2
1 1 2 1
d7 1 d 6 1 d7
2 2
4 4 2
2 1
Taking the smaller root gives d 7 0.5000
2
1
d8 1 d7 1 0 d8 1 (as expected)
2 2
4
By symmetry d 0 d8 1
d1 d7 0.5000
d 2 d6 0.292893
d3 d5 0.133975
(c) Plot the compressor characteristic for part (b) (see right column of Table above).
P.1.2 In a compact-disc (CD) digital audio system, 16-bit linear PCM is used with a sampling frequency
of 44.1 kHz for each of two stereo channels.
(a) What is the resulting data rate?
Data Rate = (2 Channels)x (44,100 symbols/sec per Channel) x (16 bits/symbol)
= 1,411,200 bit/s = 1.4112 Mbit/s
(d) If music has a loading factor = 20, find the average S/Dq in dB.
2V 2
Here the step size is given by , where N is the number of bits.
Q 2N
The average signal power is given by S 2
2
2
The average distortion is given by Dq N Q
12 3 22 N
S S 2 3 22 N
The signal-to-distortion ratio is then
Dq NQ NQ 2
With and N = 16, and = 20, this becomes
Companding
For real signals, such as speech and video, a linear quantizer (i.e uniform quantizer) is not the
optimum choice, in the sense of achieving minimum mean square error. Yet the uniform quantizer
is simple to design.
The option is to retain the linear quantizer, but transform the input signal by preceding the
quantizer with a compressor. The receiver incorporates an expander that provides the inverse
characteristic of the compressor. This technique is called companding.
Companding Technique
Logarithmic Companding
For any signal, the goal is to provide a constant S/Dq ratio over all the dynamic range of the
signal.
-Law Compandig
For = 0, there is no compression, and the result is a linear quantizer.
= 255 is used for most applications of speech processing
x
ln 1
x
F ( x) sgn( x) x
0 1
max
ln 1 xmax
A-Law Compandig
For A = 0, there is no compression, and the result is a linear quantizer.
A = 87.6 is used for most applications of speech processing
x
ln 1
xmax x
F ( x) sgn( x) 0 1
ln 1 xmax
ln 1 xmax 1 ln A x 1
sgn( x ) x 1
1 ln A A
Example 1.4: For = 255, 16-segment companded PCM, determine the code word that
represents a 5-volt signal if the encoder is designed for a 10-volt input range.
What output voltage will be observed at the PCM decoder?
What is the resulting quantizing error?
Solution
Since a 5-volt signal is half the maximum input value, the corresponding PCM amplitude is
represented by (1/2) 8159 = 4080
Insensitive to variations in input signal power and also insensitive to the actual pdf model – Both
desirable properties.
In Differential PCM, the differences between successive samples are PCM-encoded and
transmitted.
If such differences are transmitted, then simply by adding up (accumulating) these changes can
generate at the receiver a waveform identical in form to the original.
At t = kT, let the sample x(kT) be represented as x(k) = x(kT), i.e. let x(k) = x(kT)
We may well anticipate that the differences x(k) - x(k-1) will be smaller than the sample
values themselves. Hence fewer levels will be required to quantize the difference than are
required to quantize x(k) and correspondingly, fewer bits will be needed to encode the levels.
Practical Problems
Practically, there are times when the differences may turn out to be very large. In that case,
there are at least two options:
1. One option is to increase the sampling rate to make the differences smaller
(closely-spaced samples will be near in amplitude). But this necessarily increases the bit
rate, thereby requiring more bandwidth.
2. Another option is to use a predictor that uses the previous values of the differences to
predict the next value for the difference signal, and this is then quantized and encoded.
Linear Predictor
If an analogue signal x(t) is oversampled (several times the Nyquist rate), the samples exhibit
high correlation, and this can be used in designing a predictor.
The quantity to be minimised is the mean square error, and the minimisation is done with
respect to the coefficients a k , k =1, 2, 3, …, m. The mean square error (MSE) is
E x(n) xˆ(n)
2
which we can expand to give
E x 2 (n) 2Ex(n) xˆ(n) E xˆ 2 (n)
Substituting for xˆ(n) , we can write this as
m
E x (n) 2 E x(n) x(n i)ai E x(n i)ai x(n j )a j
m m
2
i 1 i 1 j 1
Equivalently,
i 1 i 1 j 1
Setting a to zero, we obtain
k
m
a r
i 1
i k i rk for k 1, 2, 3, ,m
r0 r1 rm1 a1 r1
r1 r0 a
2 r2
r0 r1
r r0 a r
m1 r1 m m
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 60
The coefficient matrix is Toeplitz matrix, which is another way of saying that the terms along
each diagonal are equal. It is so-named after the person who first published the work on
matrices of this form; (accordingly, the name begins with a capital T).
s(t) { si } { ei } { ei + q i }
Sampler Quantizer
Input + -
{ si } +
Predictor
+
Transmitter
Digital
Channel
{ si + q i } s(t) + q(t)
Low-Pass
+ Filter Output
+
Predictor
{ si }
Receiver
Differential PCM (DPCM).
a21
a11r0 r2
and a22
r 2 r0 a11
2
1 1
(c) Show that the residual error 2 is given by
2 r0 1 a11
2
1 a22
2
For the general solution, we define the three vectors, a m , rm , and rmB as:
am1 r1 rm
am 2 r rm1
am rm 2 ,
rmB
a r r
mm m 1
Example: Given a signal with the autocorrelation function r [3, 2, 1, 0.5, 0.25] , use the
Levinson-Durbin algorithm to construct a suitable linear predictor.
Below is the figure of a third-order linear predictor, in which the order of the filter has
been suppressed in the subscripts for the filter coefficients.
k1 = r1/r0 = 2/3
Xn Xn-1 Xn-2 Xn-3 = 0.667 k1 = 0.667
z-1 z-1 z-1
a11 = k1 = 0.667 1 = r0[1– (k1)2] = 1.667
a1 a2 a3 a21 = 0.800 k2 = – 0.200
X X X a22 = -0.2000 2 = 1[1 – (k2)2] = 1.600
The Table shows the predictor coefficients for the autocorrelation sequence
[3, 2, 1, 0.5, 0.25]. Since there is no significant improvement in going from order 3 to order
4, a third order predictor is sufficient: a1 = 0.7692, a2 = –0.2308, a3 = –0.0385.
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
rm1 aTmrmB
4. am1, m1
r0 aTmrm
Equations 3 and 4 and 4 are the key equations in the algorithm
Knowing am, and rm, use rm+1 in equation 4 to determine am+1,m+1 = km+1
Substitute in equation 3 to obtain
5. am1, j am, j km1am,m j 1 for j=1,2, …, m
m 2
i 1
That is
m E x n 2 ai E x n i x n ai a j E x n j x n i
m m m
2
i 1 i 1 j 1
m m m
m r0 2 ai ri ai a j ri j
i 1 i 1 j 1
rm aTm1rm1, B rm aTm1rm1, B
and am,m km
r0 aTm1rm1 m1
Therefore the MSE becomes
m 1
m r0 am1,i km am1,mi ri am,m rm
i 1
m 1
m1
m r0 am1,i ri km am1,mi ri am,m rm ( recall am,m km )
i 1 i 1
m 1
m 1
m r0 am1,i ri km rm am1,mi ri
i 1 i 1
The first term and the second term in round brackets are, respectively
m 1
m 1
r0 am 1, i ri m1 and rm am1,mi ri km m1
i 1 i 1
r1
Now 0 r0 and k1
r0
Since the MSE must decrease as the order of the prediction filter is increased, it is necessary
the magnitude of the “reflection coefficients” kj be less than unity; that is k j 1 for j = 1, 2,
Autocorrelation Values for a Signal r k
k r k k r k 6.0 5.75
5.0
0 5.75 10 2.10
1 3.70 11 1.85 4.0 3.70
2 0.80 12 0.90 3.0
2.10
3 -1.90 13 -0.30 2.0 2.00 1.85
4 -3.15 14 -1.50 1.0 0.80 0.70 0.90 0.55
0.25 0.15
5 -3.35 15 -1.40 0.0
-0.30 -0.45
6 -2.70 16 -0.45 -1.0 -1.10
7 -1.10 17 0.25 -2.0 -1.90 -1.50 -1.40
8 0.70 18 0.55 -2.70
-3.0 -3.15
9 2.00 19 0.15
-4.0 -3.35 k
0 2 4 6 8 10 12 14 16 18 20
Solution:
k1 = r1/r0 = 3.70/5.75 = 0.64348
1 = r0 [1 – (a11)2] = 5.75[1 – 0.643482] = 3.36913
a11 = k1 = 0.64348
--------
k2 = (r2 – a11* r1) /1 = (0.80 – 0.64348*3.70)/3.36913 = -0.4692
2 = 1[1 – (k2)2] = 3.36913 [1 – (-0.4692)2] = 2.6274
a21 = a11 – k2*a11 = a11(1– k2) = 0.64348 (1–(-0.4692)) = 0.9454
a22 = k2 = -0.4692
--------
k3 = (r3 – a21*r2– a22*r1) /2
= (-1.90– 0.9454 *0.80 –(-0.4692)* 3.70)/ 2.6274 = –0.35026
3 = 2[1 – (k3)2] = 2.6274 [1 – (–0.35026)2] = 2.30507
Delta Modulation
When the sampling rate is sufficiently high, the DPCM samples can be represented by just one
bit, indicating “up or down” (above or below), ( ) and the DPCM then termed delta modulation.
Step size
TS Sample
spacing
Transmitted
Sequence 1 1 1 1 1 0 0 1 0
The low-pass filter in the receiver smoothes the staircase signal to recover the analog signal.
If the slope of the input signal exceeds the slope of the staircase (/T), the resulting error is
known as slope overload distortion.
For an input slope less than /T, the errors are a form of quantizing distortion known as
granular noise.
The step size is chosen to minimize slope overload by continuing to increase the step size for
runs of three consecutive like bits, and to decrease the step size when there is alternation in
bit patterns
Dump Switch,
SW1
C Sample Switch,
V
SW2
s(t) v0(t)
+ v0(T)
0 t R Sampled at
T
t =T
T T
1 1 VT
s0 (T ) s(t )dt Vdt
0 0
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 71
The sample voltage due to the noise is
T
1
n0 (T ) n(t )dt
0
The sampled noise voltage n0(T) is a random variable as opposed to n(t), which is a stochastic
process (i.e., a random process). The variance of n0(T) is found as follows:
02 E n02 (T )
1
2 0
T
T
0
E n(t1 )n(t2 ) dt1dt2
1
2
T
0
T
0
N0
2
t1 t2 dt1dt2
where we have used the fact that the noise is white. Finally,
N 0T
02
2 2
We would like the output signal voltage to be as large as possible in comparison with the noise
voltage. Hence a figure of merit of Interest is the signal-to-noise ratio
s0 (T )2
2 2 2E
V T S
E
n02 (T ) N0 N0
(a) The signal output and (b) the noise output of the integrator
Note that the signal-to-noise ratio increases with increasing bit duration T and that it
depends on V2T which is the normalized bit energy of the signal.
Therefore, a bit represented by a narrow, high amplitude signal and one by a wide, low
amplitude signal are equally effective, provided V2T is kept constant.
PROBABILITY OF ERROR
The error probability Pe for the integrate-and-dump receiver above is found as follows:
We assume that the sampled noise n0(T) is Gaussian. That is,
1 x2
f S0 ( x ) exp
2 2
2 0
2
0
where 02 , is the noise variance (already determined above). An error will occur if either the
signal transmitted was positive and the noise component less than -VT/ or when the
transmitted signal is negative and the noise level is higher than +VT/ . The latter event has
probability represented by the area under the tail of the pdf curve as shown
x2
1
Pe f S0 ( x)dx exp dx
VT / VT / 2 2
2 02 0
Defining
u x/ 0 2 , the expression for P may be rewritten as
e
1 2
Pe
1
exp u 2 du erfc V 2T/N0
2 2
u V 2T/N0
Noting that ES = V2T is the signal energy of a bit, we have V 2T/N0 Es / N0 . So we can write
Pe as
1
Pe erfc
2
Es /N0
where, the complementary error function erfc(x) is defined as
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 73
erfc x
2
x
exp u 2 du
Schwarz’s Inequality
2
2
If we wish to distinguish between s1(t) and s2(t), then we form the difference signal
p(t) = s1(t) – s2(t),
If the input signal to the system is this signal p(t), then the output will be
p0(t)= s01(t)–s02(t),
Let the respective Fourier Transforms be P(f) and P0(f). If H(f) is the transfer function of the
filter, then P0(f) = H(f)P(f), and
0 H ( f )P( f )e df
j 2fT j 2fT
p0 (T ) P ( f ) e df
The output noise n0(t) has power spectral density given by
2
Gn0 ( f ) H ( f ) Gn ( f )
By Parseval‟s Theorem,
2
02 Gn0 ( f )df H ( f ) Gn ( f )df
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 74
The signal-to-noise ratio is then
2
H ( f )P( f )e df
j 2 fT
p02 (T )
02
H ( f ) G ( f )df
2
n
p02 (T ) P( f ) 2
2
Y ( f ) df df
02 Gn ( f )
When the equal sign applies, we have the optimum filter as transfer function as
P * ( f ) j 2 fT
H( f ) K e
Gn ( f )
The impulse response of this filter is then obtained as the inverse Fourier transform
A physically realisable filter will have a real impulse response h(t) = h*(t), and so we must have
2K
h(t ) p(T t )
N0
This is the filter that is matched to the pulse p(t). In our example, we started with
p(t) = s1(t)–s2(t). So in this case the matched filter impulse response is
h(t )
2K
s1 (T t ) s2 (T t )
N0
The signals
(a) .s1(t).
h(t) = p(T–t),
0 T/2 T t t
0 T/2 T
s2(t) h2(t)= s2(T- t)
t t
0 T/2 T 0 T/2 T
s3(t) h3(t)= s3(T- t)
t t
0 T/2 T 0 T/2 T
s4(t) h4(t)= s4(1.5T- t)
0 T/2 T 3T/2 t t
0 T/2 T 3T/2
p02 (T ) 2
2
2 P( f ) df
0 max N 0
At this point, we use Parseval’s Theorem:
2 2
P( f ) df p(t ) dt (Parseval‟s Theorem)
and also noting that p(t) = s1(t) – s2(t) for 0 t T, we write
p02 (T ) 2 T
2
2 s1 (t ) s 2 (t ) dt
0 max N 0 0
p 02 (T ) 2 T 2 T T 2
2 0 1
0 max N 0
s
(t ) dt 2
0
s1 (
t ) s 2 (t ) dt
0
s 2 (t )dt
2
E S1 2E S12 E S 2
N0
where ES1 and ES2 are respectively, the energies in s1 and s2, while ES12 is the energy due to the
correlation between s1(t) and s2(t). We define the correlation coefficient between the signals
as
T
1
12 s1 (t ) s 2 (t )dt .
E S1 E S 2 0
When the signals have the same energy ES1 = ES2 = ES, we can write the maximum signal-to-
noise ratio as
p02 (T )
S 1 12
4E
2
0 max N0
where 1 12 1 .
p02 (T ) 8E
2 S
0 max N0
Thus binary antipodal signalling is 3 dB better than orthogonal signalling (double SNR).
1.2
1 .0
1.0
0 .8
0.8
0 .6
0.6
0 .4
0.4
0 .2
0.2
0 .0 0.0
0.0 0 0.2 5 0.5 0 0 .7 5 1.0 0 1 .2 5 1 .50 0.00 0 .2 5 0.50 0 .7 5 1.00 1 .2 5 1.50
t im e , t tim e , t
p o (t)
3.6
3.2
2.8
2.4
2.0
1.6
1.2
0.8
0.4
0.0
0.00 0.25 0.50 0 .7 5 1 .0 0 1.25 1.50 1.75 2.00 2.25 2 .5 0
Nonreturn-to-Zero (NRZ)
With nonreturn-to-zero (NRZ), the signal level is held constant at one of two voltages for the
duration of the bit interval. If the two allowed voltages are 0 and V, the NRZ waveform is said
to be unipolar, because it has only one polarity. This signal has a nonzero dc component at one-
half the positive voltage, assuming equally likely 1 's and 0's. A polar NRZ signal uses two
polarities, V, and thus provides a zero dc component.
NRZ(M) – a level change is used to indicate a mark (that is, a 1) and no level change for a space
(that is, a 0);
NRZ(S) – similar except that the level change is used to indicate a space or zero. Both of these
formats are examples of the general class NRZ(I), also called conditioned NRZ, in which level
inversion is used to indicate one kind of binary digit.
Advantage of diphase.
From the diphase waveforms shown in the figure, it is readily apparent that the transition
density is increased over NRZ(L), thus providing improved timing recovery at the receiver, and
this a significant advantage of diphase.
Binary 0's are always coded as level 0; binary 1's are coded as +V or -V where the polarity
alternates with every occurrence of a 1. Bipolar coding results in a zero dc component, a
desirable condition for baseband transmission.
As shown in the following figure, bipolar representations may be NRZ (100 percent duty cycle)
or RZ (50 percent duty cycle).
Bandwidth
o The bandwidth available in a transmission channel is described by its frequency
response, which typically indicates limits at the high or low end.
Spectrum shaping
o Spectrum shaping can help minimize interference from other signals or noise.
o Conversely, shaping of the signal spectrum can allow other signals to be added
above or below the signal bandwidth.
RX() = E { x(t)x(t + ) }
FEE 422: Telecommunications and Electro-Acoustics B V. K. Oduol Page - 83
where E{} represents the expected value or mean. The power spectral density describes the
distribution of power versus frequency and is given by the Fourier transform of the
autocorrelation function (Wiener-Khintchin Theorem)
R X ( )e d
j 2f
SX ( f ) (Wiener-Khintchin Theorem)
The following figure indicates a coder/decoder block diagram and waveforms for bipolar signals.
The bipolar signal is generated from NRZ by use of a 1-bit counter that controls the AND gates
to enforce the alternate polarity rule. Recovery of NRZ(L) from bipolar is accomplished by
simple full-wave rectification.
An error detection capability results from the property of alternate mark inversion.
Consecutive positive amplitudes without an intervening negative amplitude (and vice versa)
are a bipolar violation and indicate that a transmission error has occurred.
The advantages of bipolar transmission have made it a popular choice, for example, by
AT&T for T1 carrier systems that use 50 percent duty cycle bipolar.
Solution:
Replace the string of 0's with a special sequence that contains intentional bipolar
violations, to create additional transitions that improve timing recovery.
This "filling" sequence must be recognized and replaced by the original string of zeros at
the receiver.
A commonly used bipolar coding scheme for eliminating strings of zeros is Bipolar N-
Zero Substitution (BNZS)
Bipolar N-Zero Substitution (BNZS)
Bipolar N-Zero Substitution (BNZS), replaces all strings of N 0's with a special N-bit sequence
containing at least one bipolar violation.
All BNZS formats are dc free and retain the balanced feature of AMI, which is achieved by
forcing the substitution patterns to have an equal number of positive and negative pulses.
These substitution rules can also be described by use of the following notation:
B represents a normal bipolar pulse that conforms to the alternating polarity rule,
V represents a bipolar violation, and
0 represents no pulse.
B3ZS
Thus, in the B3ZS code, each block of three consecutive 0's is replaced by BOV or 00V; the
choice of B0V or 00V is made so that the number of B pulses (that is, 1 's) between consecutive
V pulses is odd.
B6ZS
B6ZS replaces six consecutive 0's with the sequence OVB0VB.