Waveform Coding (PCM and DM) 95
2T 3T 4T 8T 9T 10T 11T
t
Sample Number S, S, S. s, s, S. S
Actual values of
1.7 4.4 7.2 7.4 73
the samples 7.0 2.4 0.4 3.4 5-3 4-3 3-2
Rounded off
2 4 7 7 7 7 2
values 3 5 4
Binary code for O10 100 111 111 111 111 010
rounded value 000 O11 101 100 O11
Fig. 3.27 llustrating the quantization process
Definition of Quantization It is the process of assigning to each one of the sample values of
Signal, a discrete value fromaprescribed set of a finite number of such discrete the message
values, called the `quantized
values".
We note that for any given dynamic range of the analog signal, if the number of
the 'quantized levels', is increased, the interval between two the 'prescribed levels', or
Smaller and so the error due to quantization, which can be at successive levels, called the 'step size', becomes
the most + 0.5 (step size), also becomes smaller.
Ihis will certainly give a better approximation to the original analog
Detter approximation by increasing the quantization levels, we have tomessage. However, in order to get a
pay aprice in the form of increased
bandwidth. as we are going to see later.
94 DigitalCommunication
It iseasy to find the impulse response of a zero-order-hold. If a unit h(t)|
impulse ö) is given as input to the system,
y(0) = 6() - 6(t- T)
z() = y() dt =u() - u(t -T)=p) ...(3.32)
where, p() = impulse response h() and is as shown in Fig. 3.26.
The transfer function H) is therefore given by
...(3.33)) T,
H) = FÊh()] = T, sinc (fT) en,
Hence, the output of the ZOH for an inputof Fig. 3.26 Impulse response of ZOH
xg() =k=-oo x(KT) S( - kT,)
is a staircase waveform as shown in Fig. 3.25 (b). This contains several high frequency components outside
the baseband. Hence,the ZOH is generally followed by an LPF with a cutoff frequency of W. To compensate
for the aperture effect, an amplitude equalizer of appropriate transfer function, as discussed earlier, will aso
be added in tandem with the ZOH and LPF.
From Eq. (3.33)which gives the transfer function of the ZOH, two things are quite clear.
(i) ZOH gives a linear phase shift corresponding to a time delay of T/2.
(ii) Since the spectrum of the reconstructed signal is equal to X ).H)for Ifls W, and since H() isa
sinc functionwhile X; f) = X) for Ifl < W when f, >2W, it followsthat the reconstructed signal isa
distorted version of x(). As mentioned earlier, we make use of an amplitude equalizer to reduce this
distortion.
3.2.6 Uniform Quantization
A continuous-time signal is defined for all values of time and its amplitude can have any value within the
limits set by its dynamic range. When we sample it, what we get is only a discrete time signal and not adigital
signal. Sampling, as we have seen, discretizes the continuous time signal only in time but not in amplitude
The samples obtained by the sampling process can have a continuum of values they are not restricted to
any finite set of prescribed values. So, the next step in the digitization of an analog signal is the discretization
of the amplitudes of these samples obtained through the sampling process. For this, let us say we divide the
dynamic range of the analog signal into acertain finite number of equal segments, as shown in Fig. 3.27. We
round off asample value falling within aparticular segment to the value represented by the 'prescribed level
passing through the middle of that segment. This process of rounding off is called guantization.
Quantization certainly introduces errors into the values of the samples. This, in turn, may be viewed a
deliberately distortingthe original message signal. However, we now have samples whose values are restricted
to acertain finite set of prescribed values. This means that we have discretized the amplitudes of the sampl
Thus, the original message signal has been discretized in time by the sampling process and in amplitude
the quantization process. Since the permitted, or prescribed levels are equidistant, this quantization is Ca
*uniform quantization'. It must be noted that the quantization process which we are adopting Is
Sucal
it is instantaneous or memoryless in the sense that the way the rounding off of a sample is done is solely
dependent upon its actual value only and is not in any way influenced by the values of the other sa
preceding it or succeeding it. Before proceeding to the quantizers, their types andcharacteristics, let us
'quantization'.
96 Digital Communication
3.247 Quantizers
signal to a 'quantizer", which transfor
Quantization can be performed by feeding the samples of the analog
each of the samples fed to it into a'quantized sample' having an amplitude coresponding to the 'prescrib
within the pertinent interval (step) into which
level' used for representing any sample value falling Figures 3.28(a) and (b).
analog signal sample falls. This is shown diagrammatically in
sample
signal Xm+2
Am+i
analog Xam+1 XE Ami
Quantizer
Q
Xm+1 i=0,1, 2, ...
Xam+1
the
of am Am
value
X=
xE Ami
i= 0, 1, 2, Xami
(a) (b)
Fig. 3.28 (a) and (b) Illustrating the action of a 'quantizer'.
As shown in these figures, any sample value xfalling in the interval x, to x,pj; i.e., in the interval (or step)
A, willbe mapped by the quantizer into the prescribed value x corresponding to that interval.
As stated earlier, a quantization process in which the quantization levels are uniformly spaced is called
"uniform quantization', and a quantizer which performs uniform quantization is called a uniform quantizer.
A quantization process in which the quantization levels are not uniformly spaced. is called non-unifor
quantization and the quantizer performing non-uniform quantization, is calleda 'non-uniform quantizer. e
shall presently discuss only the uniform quantizers.
Uniform Quantizers These are of two types depending on the shape of
(i) Mid-tread type
their characteristics.
(ii) Mid-rise type
As shown in these figures, the quantizer characteristics have a
a mid-tread type or a mid-rise
type. The only
staircase shape irespective of whetheritinis
the 'tread' portion and that is why it difference that in the mid-tread characteristic, the originis
is
is called mid-tread;
portion of the staircase type of whereas for the other one, the origin is in the 'rise
the mid-tread type, any input characteristic
value
and so it is named the mid-rise type of quantizer. Note that in
value between 0.5 and 1.5 is mappedbetween
to
-0.5 and +0.5 is mapped to an
output value of 'zero': any
input
hand, any input value between 0and I an output value of1, and so on, In the the other
is mapped to an is mappedto an mid-rise
quantizer, con l and2
output value of 1.5, and so on. output value of 0.5: any input value between
Waveform Coding (PCM and DM) 97
Output
Output
4
3 3-54
2 2.5
1 1.5
-4 -3-2 -10.5
T-43 2 1oo.51 2 3 4 -0.5 1 2 3 4
t-1
Input Input
-1.5
2.5
-3
-4 t-3.5
(a)
(b)
Fig. 3.29 Quantization characteristic: (a) Mid-tread Quantizer (b) Mid-rise Quantizer
Example 3-7
The temperature at a particular place varies between 14°C and 34°C. For the purpose
of transmitting the temperature record of that place using PCM, the record is sampled at an appropriate
sampling rate and the samples are quantized. If the error in representation of the samples due to
quantization is not to exceed +1% of the dynamic range, what is the minimum number of quantization
levels that can be used?
IfQ is the number of levels and Ais the step size,
(34 14)
A=
Quantization error =+in
2 the representation of any sample
100×=l or
0= 20 X_S0
2
:. the minimum number of 'prescribed levels', or 'quantization levels' that can be used is 50.
3.2.8 Quantization Noise
Let the message signal x(t) be azero-mean random process. Then its sample, x, is a zero-mean random
Variable. If this sample, x, is fed to a quantizer, and if x, is the prescribed quantization level to whichxis
approximated, the quantization error e, is given by
98 DigitalCommunication
Now, since in a quantizer, the levels x,'s are symmetrically located on either side of zero level. and
xis a zero-mean random variable, the random variable e, must also be zero-mean. Since x, is at the cen Sul
the interval of sample amplitudes into which, the sample x, has fallen, we also knowthat
...334
where, Ais the 'interval' of amplitudes, or the 'step size'. That is, the error may take any value from-4 3.2.
+52 However, the way the error random variable, e, is distributed over this range of values, is not know Cons
peak
But, if the step size, A, is small, i.e., if the number of prescribed quantization levels, , is large, it is quie sinu
A
reasonable to assume that the random variable, e, is uniformly distributed over the interval to+
we shall assume that e, has a PDF as shown in the following figure.
and
-A/2 A(2 pe
Eo
is
Fig. 3-3o PDF of the quantization error r.v.
These random errors caused by quantization in the successive samples, appear as fo
noise, called the
quantization noise. The mean-squared value of this noise can be evaluated, using the probabilistic model we
have developed for the quantization-error random variable, e.. a
Mean-squared value of the error = average power in the quantization noise
+oo
,=leflede,
But f. (e), the probability density function of e is
t
f. (e) = MA;-se,s5
0; otherwise
-A2
. .(3.35)
The step size, of course,
of depends upon the dynamic range of the number
prescribed quantization levels O. So, let us message signal x(t) and the1.Then the
dynamic range of values of x(t), is -1 to +1. So,normalize the message signal x() so that Lx(t)l s
we may write
A= (since Q>> 1) ...(3.36)
Waveform Coding (PCM and DM) 99
Substituting this for Ain the expression for e, we get
1
average power in quantization noise = e=
(x(0)l s 1) 30 ...3.37)
3,2.9 Non-uniform Quantization (Companding)
Consider a uniform quantization with characteristics as shown in Fig. 3.29 for which an input of say 8volts
peak-to-peak constitutes full load. Let us first consider the case when it is given the full 8 volts peak-to-peak
sinusoidal signal as input. Then the signal power is
s==8w
and the noise power is
N, 12 12 w (see Eq.3.35)
Hence, the (SNR),-(==96 = 19,8 dB
But suppose, to the same quantizer whose full load input is 8 volts peak-to-peak, we give a 2 volts peak to
peak sinusoidal signal as input. Then the signal power is 0.5 W but quantization noise power, which, as per
Eq.(3.35) depends only on A, the step-size, remains at the same value of (1/12) watt. Thus, the (SNR),now
is 6, which is equal to 7.78 dB.
Thus, we find that a uniform quantizer gives good (SNR), for large-amplitude signals but a poor (SNR),
for low-amplitude signals. This is not desirable. A quantizer which gives almost the same (SNR), for signals
whose powers vary over a wide range, is called a robust quantizer. Aswe will be seeing presently, employing
a non-uniform quantization is the way to realize robust quantizer.
Thus, uniform quantization is quite suitable for message signals which have an amplitude probability
distribution that is uniform over their dynamic range. But, unfortunately, in the case of speech, which is perhaps
the most important message signal in communication engineering, the amplitude probability distribution is
not at all uniform. Since the ratio of intensities corresponding to a loud shout and a whisper can be as high as
1000:1, the dynamic range of aspeech signal is quite large. However, a typical speech signal is characterized
by a lowamplitude for most partof the time, corresponding to normal speech, with occasional peaks of large
amplitude but of short duration, corresponding to shouts. Thus, while its dynamic range is large, its average
power is quite small. So, for speech signals, if we wish to use only uniform quantization, we are left with two
alternatives, both of which are unattractive:
i) Keep the quantization error and quantization noise low by using a large number of
quantization
levels. i.e.. have a large Q. As mentioned earlier, and as would be seen later when we discuss PCM
in detail, if Ois large, the transmission bandwidth would be correspondingly large.
(ii) Keep the bandwidth requirement low by using a smallnumber of prescribed quantization levels,
i.e.,
a low Q. in which case, the quantization noise will be high.
Non-uniform quantization provides a good solution for the above problem, as it allows us to keep the
quantization noise low even when is small, i.e., even when we use a small number of prescribedquantization
levels.
As noted earlier, the mean-squared value of quantization noise is
Size. Further, for speech signals, the signal amplitude is low for proportional to the square of the step
most part of the time and large peaks occur
only very rarely and they are of very short duration. Hence, in the case of
speech signals, in order to obtain
100 Digital Communication
a good signal-to-noise ratio, it makes sense to provide smaller step size in the low amplitude region,
dynamic range of thesignal. This means that yWe
large step size in the large amplitude region of the quantization levels are not uniformly space
employ non-uniform quantization, wherein the prescribed message speech sion
Non-uniform quantization is equivalent to first subjecting the samples of the
then applying uniform quantizatio
amplitude compression by passing them through a compressor and
samples is accomplished according to a specifie
these compressed samples. Compression of the input input and output samples. There are two diffe
governing the relationship between amplitudes of the
compression laws in vogue and they are
(ii) The A-lawod ogt l e e
(i) The u-law
output amplitude and the input amplitude f
Both these laws provide a near linear relation between the
the input samples, and a somewhat logarithmic relation for the larger amplitude in
small amplitudes of
samples.
The u-law If xand y are respectively the nornalized
input and output values, this compression law
defined by the following input-output relation:
log[1 + u lxl]
lyl =
log[1+]
characteristics of a u-law compressor are given
where, uis apositive constant. The typical input-ouiput noted that when u=0, it corresponds to
in Fig. 3.31 for three different values of the constant u. It may be
is higher for larger values ofu.
uniform quantization and that compression of the larger sample values
1.0
=100
=10
O.8
LL = 0
o.6 Y,
Y4
0.4
O.2
o.6 0.8 1.0
0.4 X,
Fig. 3-31 -law characteristic
Waveform Coding (PCM and DM) 101
The A-law The other compression law in vogue, is the A-law which is defined by the following input
output relation:
Abd
0slils
1+ logA A
lyl = ...(3.39)
1+ log[All]
1+ logA A
where, Aisa constant. Figure 3.32 shows the input-output relation for an A-law compressor. Clearly, A = 1
corresponds to uniform quantization. Larger the value of the constant A, more is the compression.
In both -law as well as in A-law, as uor A increases, the compression for the larger amplitudes also
increases and so the dynamic range increases. Again, in both the cases, the signal-to-quantization noise ratio
becomes poorer and poorer for larger amplitudes (because of larger step size) as the or Ais increases.
Hence, the choice of the value of uor A is a matter of compromise and in practice u = 255 and A = 87.6 are
generally used.
1.0
100
0.8
A=1
0.6
0.4
0.2
O.2 0.4 O.6 0.8 1.0
Fig. 3-32 A-law characteristics
Since the message signal is compressed at the transmitter, in the receiver it has to be expanded to remove
the distortion. For this purpose, an expander haVing an input-output characteristic which is the exact inverse
of that of the compressor at the transmitter, will have to be used. This process ofcompressing message signals
like speech, at the transmitter and expanding them at the receiver, is called 'companding'. Companding is
aword coined by combining the words, 'compressing' and 'expanding". Figure 3.33 shows the compressor
and expander characteristics together, for both positive as well as negative values of input samples.
102 Digital Communication
Compressor
characteristic Expander
characteristic
st
We
inp
Sin
the
input
both positive and negative values of
Fig. 3-33 Compressor and expander characteristics for ste
the normalized inp
that in u-law as well as A-law, the input represented by x is
Note: It must be noted also the normalized output.
Correspondingly. the output represented by y., is equivalent
quantization of the message x may be considered can easily t
Earlier we had stated that non-uniform quantization. This
compressed signal, y, to uniform
compressing x and then subjecting the -law compressor characteristic. x is the message input
to Us
understood from Fig. 3.31 which shows the unito
figure, the compressor output is subjected to
compressor and y is the compressor output. In that quantization levels.
yg being the uniformly spaced
quantization, the dotted lines passing through y;- Y2x -. levels e
passing through y): y2: .... J'g are equivalenta
concerned, these quantization levels
so far as the input, x, is Clearly, these prescribed quantization
input-side quantization levels passing through x, Xz, .... Xg. input values, making the step siZe S
the input are closer for smaller values and farther apart for larger
for smaller values of the input and large for larger values of input.
Example3.8 with Q
A message signal Am cos ,t, is sampled and applied to a quantizer outputof
loads Find the signal-to-quantization noise ratio at the
permitted quantization levels, fully it.
the quantizer.
o
to+4,
Since the signal fully loads the quantizer, it means that the Qlevels cover the full range of - A,m
the signal.
t
Hence, step size 2A,m
. quantization noise mean-squared value = A
12N,
