EE05321Notes 8

Pulse Code Modulation
topics covered
¾ Principle of Pulse Code Modulation;

¾ Signal to Quantization Noise Ratio for uniform quantizaton;
A schematic diagram for Pulse Code Modulation is shown in Fig 3.11.1. The
analog voice input is assumed to have zero mean and suitable variance such that the
signal samples at the input of A/D converter lie satisfactorily within the permitted single
range. As discussed earlier, the signal is band limited to 3.4 KHz by the low pass filter.
k= Instance
Analog Speech x(t) x (kTs) xq(kTs) of sample
L Sample Encoder Serial
Analog P & & Output
Voice F Hold Quantizer P/S
t Rb
x(t) Error →Noise due bits/sec
to quantization
Ts
xn
(t ) LPF Decoder
x’q(kTs)
Fig. 3.11.1 Schematic diagram of a PCM coder – decoder
Let x (t) denote the filtered telephone-grade speech signal to be coded. The
process of analog to digital conversion primarily involves three operations: (a) Sampling
of x (t), (b) Quantization (i.e. approximation) of the discrete time samples, x (kTs) and (c)
Suitable encoding of the quantized time samples xq (kTs). Ts indicates the sampling
interval where Rs = 1/Ts is the sampling rate (samples /sec). A standard sampling rate for
speech signal, band limited to 3.4 kHz, is 8 Kilo-samples per second (Ts = 125μ sec),
thus, obeying Nyquist’s sampling theorem. We assume instantaneous sampling for our
discussion. The encoder in Fig 3.11.1 generates a group of bits representing one
quantized sample. A parallel–to–serial (P/S) converter is optionally used if a serial bit
stream is desired at the output of the PCM coder. The PCM coded bit stream may be
taken for further digital signal processing and modulation for the purpose of transmission.
The PCM decoder at the receiver expects a serial or parallel bit-stream at its input
so that it can decode the respective groups of bits (as per the encoding operation) to
generate quantized sample sequence [x'q (kTs)]. Following Nyquist’s sampling theorem
for band limited signals, the low pass filter produces a close replica x̂ ( t ) of the original
speech signal x (t).
If we consider the process of sampling to be ideal (i.e. instantaneous sampling)

and if we assume that the same bit-stream as generated by PCM encoder is available at
PCM decoder, we should still expect x̂ ( t ) to be somewhat different from x (t). This is
solely because of the process of quantization. As indicated, quantization is an
approximation process and thus, causes some distortion in the reconstructed analog
signal. We say that quantization contributes to “noise”. The issue of quantization noise,
its characterization and techniques for restricting it within an acceptable level are of
importance in the design of high quality signal coding and transmission system. We focus
a bit more on a performance metric called SQNR (Signal to Quantization Noise power
Ratio) for a PCM codec. For simplicity, we consider uniform quantization process. The
input-output characteristic for a uniform quantizer is shown in Fig 3.11.2(a). The input
signal range (± V) of the quantizer has been divided in eight equal intervals. The width of
each interval, δ, is known as the step size. While the amplitude of a time sample x (kTs)
may be any real number between +V and –V, the quantizer presents only one of the
allowed eight values (±δ, ±3δ/2, …) depending on the proximity of x (kTs) to these
levels.
Fig 3.11.2(a) Linear or uniform quantizer
The quantizer of Fig 3.11.2(a) is known as “mid-riser” type. For such a mid-riser
quantizer, a slightly positive and a slightly negative values of the input signal will have
different levels at output. This may be a problem when the speech signal is not present
but small noise is present at the input of the quantizer. To avoid such a random
fluctuation at the output of the quantizer, the “mid-tread” type uniform quantizer [Fig
3.11.2(b)] may be used.
Fig 3.11.2(b) Mid-tread type uniform quantizer characteristics
SQNR for uniform quantizer

In Fig.3.11.1 x(kTs) represents a discrete time (t = kTs) continuous amplitude
sample of x(t) and xq(kTs) represents the corresponding quantized discrete amplitude
value. Let ek represents the error in quantization of the kth sample i.e.
ek = xq ( kTs ) − x ( kTs ) 3.11.1
Let,
M = Number of permissible levels at the quantizer output.
N = Number of bits used to represent each sample.
±V = Permissible range of the input signal x (t).
Hence,
M=2N and,
M.δ ≅ 2.V [Considering large M and a mid-riser type quantizer]
Let us consider a small amplitude interval dx such that the probability density
function (pdf) of x(t) within this interval is p(x). So, p(x)dx is the probability that x(t) lies
dx dx
in the range ( x − ) and ( x + ) . Now, an expression for the mean square quantization
2 2
error e 2 can be written as:
x1 +δ / 2 x2 +δ / 2
∫ ∫
2 2
e = p( x)( x − x1 ) dx + p( x)( x − x2 ) 2 dx + .... 3.11.2
x1 −δ / 2 x2 −δ / 2
For large M and small δ we may fairly assume that p(x) is constant within an interval, i.e.
p(x) = p1 in the 1st interval, p(x) = p2 in the 2nd interval, …., p(x) = pk in the kth interval.
Therefore, the previous equation can be written as

δ /2
e = ( p1 + p2 + .....) ∫
2
y 2 dy
δ
− /2
Where, y = x-xk for all ‘k’.

So,
δ3
e 2 = ( p1 + p2 + .....)
12
δ2
= [( p1 + p2 + .....)δ ]
12
Now, note that (p1 + p2 + … + pk + …)δ = 1.0
δ2
∴ e2 =
12
The above mean square error represents power associated with the random error
signal. For convenience, we will also indicate it as NQ.
Calculation of Signal Power (Si)

After getting an estimate of quantization noise power as above, we now have to
find the signal power. In general, the signal power can be assessed if the signal statistics
(such as the amplitude distribution probability) is known. The power associated with x (t)
can be expressed as
+V
S i = x ( t ) = ∫ x ( t ) p ( x ) dx
2 2
−V
where p (x) is the pdf of x (t). In absence of any specific amplitude distribution it is
common to assume that the amplitude of signal x (t) is uniformly distributed between ±V.
In this case, it is easy to see that
⎡ x3 ⎤ V 2 ( M δ )
+V 2
+V
1
S i = x ( t ) = ∫ x ( t ) 2V dx = ⎢
2 2
= =
⎥ 3 12
−V ⎣ 3.2V ⎦ −V
Now the SNR can be expressed as,
(Mδ )
2
2
V
Si = 3 = 12 =M
2
NQ δ 2 δ
2
12 12
It may be noted from the above expression that this ratio can be increased by increasing
the number of quantizer levels N.
Also note that Si is the power of x (t) at input of the sampler and hence, may not
represent the SQNR at the output of the low pass filter in PCM decoder. However, for
large N, small δ and ideal and smooth filtering (e.g. Nyquist filtering) at the PCM
decoder, the power So of desired signal at the output of the PCM decoder can be assumed
to be almost the same as Si i.e.,
So Si
With this justification the SQNR at the output of a PCM codec, can be expressed as,
M (2 ) 4
SQNR = So 2 = N 2 = N
NQ
So ⎛ ⎞
and in dB, = 10log ⎜ S o ⎟ 6.02 NdB
10 ⎜ ⎟
NQ ⎝ NQ⎠
dB
A few observations
(a) Note that if actual signal excursion range is less than ± V, So / No < 6.02NdB.
(b) If one quantized sample is represented by 8 bits after encoding i.e., N = 8,

SQNR 48dB .
(c) If the amplitude distribution of x (t) is not uniform, then the above expression
may not be applicable.

EE05321Notes 8

Uploaded by

Copyright:

Available Formats

EE05321Notes 8

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EE05321Notes 8

Uploaded by

Copyright:

Available Formats

Pulse Code Modulation

¾ Principle of Pulse Code Modulation;

Fig. 3.11.1 Schematic diagram of a PCM coder – decoder

If we consider the process of sampling to be ideal (i.e. instantaneous sampling)

Fig 3.11.2(a) Linear or uniform quantizer

SQNR for uniform quantizer

Therefore, the previous equation can be written as

Where, y = x-xk for all ‘k’.

Calculation of Signal Power (Si)

In this case, it is easy to see that

(b) If one quantized sample is represented by 8 bits after encoding i.e., N = 8,

You might also like