Voice Digitization

Lecture 3
 Voice Digitization
--Basic Concepts--

Voice is analog in character moves in the form of waves.

3-important wave-characteristics:
 Amplitude
 Frequency
 Phase

Why Voice Digitization?

 Ensures better quality (than analog)
 Provides higher capacity (than analog)
 Deals with longer distance (than analog)

Digitization is just a discrete electrical voltage.

The amplitude of Electrical pulses can be varied to represent
characteristics of an analog voice signal. 2
 Pulse Amplitude Modulation (PAM)

 PAM is the first step in digitizing an analog waveform.

 Establishes a set of discrete times at which the input signal waveform is
sampled.
 The sampling process is equivalent to amplitude modulation of a constant
amplitude pulse train, thus, PAM.

Figure 1: PAM

3
 Nyquist rate & Nyquist frequency
Sampling/ Nyquist Theorem:
The sampling theorem states that, the original analog signal can be
reconstructed at the receiver with minimal distortion if the sampling rate in the
system equal to or greater than twice the maximum frequency f max of the
analog signal. f ≥2f
S max

The sampling frequency, has value more than twice the maximum frequency,
is known as Nyquist frequency.

Undersampling is essentially sampling too slowly, or sampling at a

rate below the Nyquist frequency for a particular signal of interest.
Undersampling leads to aliasing and the original signal cannot be
properly reconstructed.

Oversampling is sampling at a rate beyond twice the highest
frequency component of the signal and is usually desired.
 Nyquist rate & Nyquist frequency

 The “Nyquist frequency” should not be confused with the “Nyquist rate”,
which is the minimum sampling rate that satisfies the Nyquist sampling
criterion for a given signal or family of signals.

 The Nyquist rate is twice the maximum component frequency of the

function being sampled.

 Thus, Nyquist rate is a property of a continuous-time signal, whereas

Nyquist frequency is a property of a discrete-time system.

5
 Pulse Amplitude Modulation (PAM)
• Spectrum of PAM Signal:

The PAM spectrum can be derived by observing that a continuous train of pulses has
a frequency spectrum consisting of discrete terms at multiples of the sampling
frequency.
The input signal amplitude modulates these terms individually. Thus a double-
sideband spectrum is produced about each of the discrete frequency terms in the
spectrum of the pulse train.

Figure 2: Spectrum of PAM Signal

6
 Pulse Amplitude Modulation (PAM)
 The original signal waveform is recovered by a low-pass filter designed to
remove all but the original signal spectrum.

 As shown in the figure 2, the reconstructive low-pass filter must have a cut-off
frequency that lies between BW and (fs – BW).

 Hence, separation is only possible if (fs – BW) is greater than BW ,i.e., (fs > 2BW).

Figure 2: Spectrum of PAM Signal

7
 Pulse Amplitude Modulation (PAM)
• Foldover Distortion:
 If the input is under sampled (i.e. fs < 2BW), the original waveform cannot be
recovered without distortion.
 As indicated in figure 3, this output portion arises because the frequency
spectrum centered about the sampling frequency overlaps the original spectrum
and cannot be separated from the original spectrum by filtering.

 Since it is a duplicate of the input spectrum “folded” back on top of the desired
spectrum that causes the distortion, this type of sampling impairment is called
“foldover distortion.” Another term for this impairment is “aliasing”.

Figure 3: Foldover spectrum produced by under sampling an input

8
 Pulse Amplitude Modulation (PAM)
• PAM System:

 Complete PAM system includes a band-limiting filter (or anti-aliasing filter)

before sampling to ensure that no source related signals get folded back into the
desired signal bandwidth.

• End-to-End PAM system:

Figure 5

9
 Pulse Amplitude Modulation (PAM)
Sample-and-hold circuit:
After filtering and sampling. the sampled level must be held constant
until the next sample occurs.
This is necessary for the ADC to have time to process the sampled value.
This sample-and hold operation results in a "stairstep" waveform that
approximates the analog input waveform.
 Quantization and Binary Coding
• Pulse amplitude modulation systems are not useful over long
distance, for the vulnerability of individual pulse amplitudes to
noise, distortion and crosstalk.

• The susceptibility of amplitude may be eliminated by converting

the PAM samples into a digital format. (Using regenerative repeaters)

• A finite number of bits are used for coding PAM samples.

• n bit number can represent 2n samples.

• The PAM sample amplitude is quantized to the nearest of a range

of discrete amplitude levels.
11
 Definition of Quantization

• A process of converting an infinite number of possibilities to a

finite number of conditions (rounding off the amplitudes of
flat-top samples to a manageable number of levels).

• In other words, quantization is a process of assigning the

analog signal samples to a pre-determined discrete levels. The
number of quantization levels, L determine the number of bits
per sample, n. L = 2n n = log 2 L

12
 Quantization

 Sampling results in a series of pulses of varying

amplitude values ranging between two limits: a min
and a max.
 The amplitude values are infinite between the two
limits.
 We need to map the infinite amplitude values onto a
finite set of known values.
 This is achieved by dividing the distance between min
and max into L zones, each of height ∆.
∆ = (max - min)/L

4.13
 Quantization Levels

 The midpoint of each zone is assigned a value from 0

to L-1 (resulting in L values)

 Each sample falling in a zone is then approximated to

the value of the midpoint.

4.14
 Quantization Zones

 Assume we have a voltage signal with amplitutes

Vmin=-20V and Vmax=+20V.
 We want to use L=8 quantization levels.
 Zone width ∆ = {20 – (-20)}/8 = 5
 The 8 zones are: -20 to -15, -15 to -10, -10 to -5, -5 to
0, 0 to +5, +5 to +10, +10 to +15, +15 to +20
 The midpoints are: -17.5, -12.5, -7.5, -2.5, 2.5, 7.5,
12.5, 17.5

4.15
 Assigning Codes to Zones
 Each zone is then assigned a binary code.
 The number of bits required to encode the zones, or
the number of bits per sample as it is commonly
referred to, is obtained as follows:
 nb = log2 L
 Given our example, nb = 3
 The 8 zone (or level) codes are therefore: 000, 001,
010, 011, 100, 101, 110, and 111
 Assigning codes to zones:
 000 will refer to zone -20 to -15
 001 to zone -15 to -10, etc.

4.16
Figure : Quantization and encoding of a sampled signal

4.17
 Types of Quantization
1. Uniform type : The levels of the quantized amplitude are uniformly spaced.
2. Non-uniform type : The levels are not uniform.

18
 Uniform Quantization
Dynamic Range:
Most ADC’s use uniform quantizers.
(-8, 8)

The quantization levels of a uniform
Output sample 7 quantizer are equally spaced apart.
XQ
5
Uniform quantizers are optimal when
3 the input distribution is uniform. When
1 all values within the Dynamic Range(DR)
-8 -6 -4 -2 -1 2 4 6 8
of the quantizer are equally likely.
Input sample X
-3 •DR is usually expressed in decibels as
-5 the ratio of the maximum amplitude
-7 signal to the minimum amplitude
Quantization Characteristic signal:
Example: Uniform n =3 bit quantizer
L=8 and XQ = {±1,±3,±5,±7}

19
Eeng 360 19
 Types of Uniform Quantization

Midtread: Origin lies in the middle of a Midrise: Origin lies in the middle of a
tread of the staircase like graph in (a), rising part of the staircase like graph (b),
utilized for odd levels utilized for even levels
20
 Nonuniform Quantization

Many signals such as speech signals have a nonuniform distribution.
 The amplitude is more likely to be close to zero than to be at higher
levels.

Nonuniform quantizers have unequally spaced levels.
• The spacing can be chosen to optimize the SNR for a particular type of
signal.
Output sample
XQ 6

2 Example: Nonuniform 3 bit quantizer

-8 -6 -4 -2 2 4 6 8

-2
Input sample
X
-4

-6

21
 Pulse Code Modulation (PCM)
• Pulse Code Modulation (PCM):

 PCM is an extension of PAM wherein each analog sample is quantized into a

discrete value for representation as a digital code-word.

 PAM system can be converted to PCM if we add ADC* at the source and DAC**
at the destination.

*ADC: analog to digital converter

**DAC: digital to analog converter

Figure 6: PCM

22
 Pulse Code Modulation (PCM)
Figure : Components of PCM encoder

4.23
 Encoding
• After quantization, a digit is assigned to each of the quantized signal levels
in such a way that each level has a one-to-one correspondence with the
set of real integers. This is called digitization of the waveform.

• Each integer is then expressed as an n-bit binary number, called code-

word, or PCM word.

• The number of code-words, M , is related to n by: 2n= M.

24
 Codeword
• Quantization followed by digitization maps input amplitudes into
PCM words.

• There are M integers, PCM words, or codewords to correspond to

the M allowed output amplitudes of the quantizer.

• Codebook is the set of all these M codewords.

25
 Analog to Digital

26
Figure 11: Process of digitization
 Quantization Noise
• The quantized signal is an approximation to the original signal and some error.

Signal, V Quantized Signal, Va

Quantization Noise, e

• The instantaneous error e = V-Va is randomly distributed within the range S/2 and
is called quantization error or noise. Where, S indicates step size of quantization.
• The mean square quantization error is S2.
• For linear quantization the probability distribution of the error is constant
within the ± (S/2).

Figure : Probability distribution of error due to linear quantization 27

 Quantization Noise
• The average quantization noise output power is given by the variance.

Where µ = mean, which is zero

for quantization noise.

• The range of quantization error ±(S/2) determines the limits of integration.

de

28
 Quantization Noise
• Signal to quantization noise ratio (SQR) is a good measure of performance of
a PCM system transmitting speech.

• If Vr is the r.m.s value of the input signal and the resistance level is 1 ohm, then
SQR is given by

29
 Quantization Noise
• If the input signal is a sinusoidal wave and Vm as the maximum amplitude, SQR
may be calculated from the full range sine wave as:

• Expressing S in terms of Vm and the number of steps, M, we have

30
 Quantization Noise
• Quantity 1.225M represents the signal to quantization noise voltage ratio for a
full range sinusoidal input voltage.
• M = 2n, where n is the number of bits used to code a quantization level.
Therefore:

• The table 1 is showing the values of SQR for different binary code word sizes for
sinusoidal input systems.
• Every additional code bits gives an increment of 6 dB in SQR.

31
 Quantization Noise
• Example: A sine wave with a 1-V maximum amplitude is to be digitized with a
minimum SQR of 30 dB. How many uniformly spaced quantization intervals are
needed, and how many bits are needed to encode each sample?

• Solution: Using Equation:

SQR = 7.78 + 20 log10 (Vm / S) Given,
The maximum size of a quantization interval is SQR = 30 dB
determined as: Vm = 1 V
S = (1) 10 –(30-7.78)/20
= 0.078 V

No. quantization intervals are needed for each polarity = 1 / 0.078 ≈ 13

Total quantization intervals = 2x 13 = 26

The number of bits required to encode each sample is determined as:

N = log2 (26) = 4.7 = 5 bits per sample

32
 Companding
• An alternative is to first pass the speech signal through a nonlinearity before
quantizing with a uniform quantizer.
• The nonlinearity causes the signal amplitude to be Compressed.
– The input to the quantizer will have a more uniform distribution.
• At the receiver, the signal is Expanded by an inverse to the nonlinearity to avoid
signal distortion. .
• The process of compressing and expanding is called Companding.

Compression + Expansion Companding

y = C ( x) x̂
x (t ) y (t ) yˆ ( t ) xˆ ( t )

x ŷ
Compress Uniform Qauntize Expand
Transmitter Channel Receiver
33
Companding

34
 Companding
• Various compression –expansion characteristics can be chosen to
implement a compandor.
• by increasing the amount of compression , we increase the dynamic
range at the expense of the signal-to-noise ratio for large amplitude
signals.
• There are two types of companding characteristics:
 µ-law Companding: used in North America and Japan
 A-law Companding: recommended by CCITT for Europe and most of the rest of
the world

[ Note: Please go through “Digital Telephony” by John Bellamy for

further understanding for different Companding techniques ]

35
 Companding

36
** Digital Signal 1 (DS1, sometimes DS-1) is a T-carrier signaling scheme devised by Bell Labs
 Companding

Figure A law and u law characteristics.

A common expression used in dealing with the “quality” of a PCM signal is signal-
to-distortion ratio (expressed in dB).
Parameters A and μ, for the respective companding laws, determine the range over
which the signal-to-distortion ratio is comparatively constant, about 26 dB. For
A-law companding, an S/D = 37.5 dB can be expected (A = 87.6) and for μ-law
37
companding, an S/D = 37 dB (μ = 225).
 Differential Pulse Code Modulation
(DPCM)
• Differential PCM:

 A special kind of PCM technique that codes the difference

between sample points to compress the digital data.

 Because audio waves propagate in predictable patterns, DPCM

predicts the next sample and codes the difference between the
prediction and the actual point.

 The differences are smaller numbers than the numerical value of

each sample on the full scale and thereby reduce the resulting
bit-stream.
38
Delta Modulation

39/45
Delta Modulation

40/45
 DM principle of operation

 The difference between the input signal sample and

the stair case approximation is quantized into only to
levels ± ∆

 If the current input signal sample is greater than the

previous sample then the DM modulator generates + ∆

 If the current input signal sample falls below the

previous sample then the modulator generates − ∆

41
 Assumptions and model for DM

We assume that:
• m(t) denotes the input message signal
• mq(t) denotes the staircase approximation
• m[n] = m(nTs), n = +/-1, +/-2 … denotes a sample of
the signal m(t) at time t=nTs, where TS is the sampling
period
• then

42/45
 Assumptions and model for DM
• we can express the basic principles of the delta
modulation in a mathematical form as follow:
e[n] = m[n] − mq [n − 1] (3.52)

error
eq = ∆ sgn(e[ n ]) (3.53)
signal
quantized
error signal

mq [n] = mq[n −1] + eq[n] (3.54)

quantized • Where e[n] is the error signal representing the difference
output
between the present sample m[n] and the latest
approximation mq[n-1] to it
• eq[n] is the quantized version of the error 43/45
 Delta Modulation

Figure: DM system. (a) Transmitter. (b) Receiver. 44/45

 Transmitter for DM System
• Comparator – computes difference between input signal and one interval
delayed version of it
n
• Quantizer – includes a hard- m q [ n ] = ∆ ∑ sgn( e[ i ])
limiter with an input-output i =1
n
relation a scaled version of
the signum function
= ∑e
i =1
q [i ] ( 3 . 55 )

• Accumulator – produces the

approximation mq[n] (final
result) at each step by
adding either +Δ or –Δ
• tracking input samples by
one step at a time

45/45
 Receiver for DM System

• Decoder – creates the sequence of positive or negative pulses

• Accumulator – creates the staircase approximation mq[n]
similar to tx side
• out-of-band noise is cut off by low-pass filter (bandwidth
equal to original message bandwidth)

46/45
 Noise in Delta Modulation Systems

Delta modulation is subjected into two quantization

errors
– slope overhead distortion
– granular noise

47/45
 Noise in Delta Modulation Systems

Slope overload distortion
• Slope over load distortion occurs when the step size is
too small for the stair case approximation to follow a
steep segment of the message signal m(t)

• Slope over load distortion can be avoided if the step size

∆ is selected according to the following equation
∆ dm ( t )
48 ≥
Ts dt
 Noise in Delta Modulation Systems

Granular noise distortion
• Granular noise occurs when the step size ∆ is too
large relative to the local slope characteristics of the
input signal
• This would cause the stair case approximation mq(t)
to hunt around a relatively flat segment of the input
waveform

49