Digital Representation of Analog

Signals
Gaurav S. Kasbekar
Dept. of Electrical Engineering
IIT Bombay
 Introduction
• Recall: digital communication systems have several advantages
over analog communication systems
❑ former have replaced or are replacing latter in most contexts, e.g.,
cellular networks, TV
• “Analog communication” and “digital communication”:
❑ in practice, all communication is via continuous signals and hence
analog in nature
❑ the message signal that is to be transmitted is either analog or digital
❑ E.g., if the source is speech, then:
o In analog communication, it is directly used to modulate a high-frequency
carrier signal
o In digital communication, it is sampled and quantized to obtain a bit stream,
which is then used to modulate a high-frequency carrier signal
• First step in digital transmission of analog source (e.g., speech,
music) is conversion of source to digital representation
• We now study:
❑ this analog to digital conversion
❑ and representation of the analog information as a sequence of pulses
 The Sampling Process
• Sampling is used to convert an analog signal to
sequence of samples that are usually spaced
uniformly in time
• Sampling rate must be chosen carefully, so that:
❑the sequence of samples uniquely defines the original
analog signal
• Sampling theorem tells us how to choose sampling
rate
• We now briefly review the sampling process and
prove the sampling theorem
 The Sampling Process (contd.)
• Consider an arbitrary signal 𝑔(𝑡) of finite energy, which is specified for all
time 𝑡
• Suppose 𝑔(𝑡) sampled at uniform rate:
❑ once every 𝑇𝑠 seconds
• Then we obtain an infinite sequence of samples spaced 𝑇𝑠 seconds apart:
❑ denoted by 𝑔 𝑛𝑇𝑠 , where 𝑛 takes on all possible integer values
• We refer to:
❑ 𝑇𝑠 as “sampling period”
❑ and 𝑓𝑠 = 1/𝑇𝑠 as “sampling rate”
• Let:
❑ 𝑔𝛿 𝑡 = σ∞
𝑛=−∞ 𝑔 𝑛𝑇𝑠 𝛿(𝑡 − 𝑛 𝑇𝑠 )
• 𝑔(𝑡) and 𝑔𝛿 𝑡 shown in fig.
• We will show that Fourier transform of sampled
signal 𝑔𝛿 𝑡 is:
1) 𝐺𝛿 𝑓 = 𝑓𝑠 σ∞
𝑚=−∞ 𝐺 𝑓 − 𝑚𝑓𝑠
❑ where 𝐺 𝑓 is Fourier transform of 𝑔(𝑡)
• 1) shows that process of uniformly sampling a
signal 𝑔(𝑡) results in a periodic spectrum with
period equal to the sampling rate
Ref: “Communication Systems” by Haykin and Moher, 5th ed
 Proof of the Claim 𝐺𝛿 𝑓 = 𝑓𝑠 σ∞
𝑚=−∞ 𝐺 𝑓 − 𝑚𝑓𝑠
• First, consider a periodic signal 𝑓𝑇0 (𝑡) of period 𝑇0
• We can represent it using Fourier series:
❑ 𝑓𝑇0 𝑡 = σ∞
𝑛=−∞ 𝑐𝑛 exp 𝑗2𝜋𝑛𝑓0 𝑡 , where
1 𝑇0 /2 1
❑ 𝑐𝑛 = ‫׬‬ 𝑓 𝑡 exp −𝑗2𝜋𝑛𝑓0 𝑡 𝑑𝑡 and 𝑓0 =
𝑇0 −𝑇0 /2 𝑇0 𝑇0

𝑇0 𝑇0
𝑓𝑇0 𝑡 , − ≤𝑡≤ ,
• Let 𝑓 𝑡 = ቐ 2 2
0, else.
❑ So 𝑓𝑇0 𝑡 = σ∞
𝑚=−∞ 𝑓(𝑡 − 𝑚𝑇0 )
• Hence, 𝑐𝑛 :
❑ 𝑓0 𝐹(𝑛𝑓0 ), where
❑ 𝐹(𝑓) is the Fourier transform of 𝑓 𝑡
• Thus:
❑ σ∞ ∞
𝑚=−∞ 𝑓 𝑡 − 𝑚𝑇0 = 𝑓0 σ𝑛=−∞ 𝐹(𝑛𝑓0 )exp 𝑗2𝜋𝑛𝑓0 𝑡
1) So Fourier transform of σ∞
𝑚=−∞ 𝑓(𝑡 − 𝑚𝑇0 ) is:
❑ 𝑓0 σ∞𝑛=−∞ 𝐹(𝑛𝑓0 )𝛿(𝑓 − 𝑛𝑓0 )
• Now, in the sampling context: 𝑔𝛿 𝑡 = σ∞𝑛=−∞ 𝑔 𝑛𝑇𝑠 𝛿(𝑡 − 𝑛 𝑇𝑠 )
• Fourier transform of 𝑔𝛿 𝑡 is 𝐺𝛿 𝑓 = 𝑓𝑠 σ∞
𝑚=−∞ 𝐺 𝑓 − 𝑚𝑓𝑠 by:
❑ Duality theorem and the fact that the 𝛿(. ) function is an even function
• Recall:
1) 𝑔𝛿 𝑡 = σ∞𝑛=−∞ 𝑔 𝑛𝑇𝑠 𝛿(𝑡 − 𝑛 𝑇𝑠 )
The Sampling Process
2) 𝐺𝛿 𝑓 = 𝑓𝑠 σ∞
𝑚=−∞ 𝐺 𝑓 − 𝑚𝑓𝑠
• Taking Fourier transforms on both sides of 1), we get:
3) 𝐺𝛿 𝑓 = σ∞
(contd.)
𝑛=−∞ 𝑔 𝑛𝑇𝑠 exp −𝑗2𝜋𝑛𝑓𝑇𝑠
❑ This relation is called:
o discrete-time Fourier transform
❑ Can be viewed as Fourier series representation of the periodic frequency function 𝐺𝛿 𝑓
• Next, suppose the signal 𝑔(𝑡) is strictly bandlimited:
❑ 𝐺 𝑓 = 0 for 𝑓 ≥ 𝑊
1
• Also, suppose we choose the sampling period 𝑇𝑠 =
2𝑊
• Then by 3), we get:
𝑛 −𝑗𝜋𝑛𝑓
4) 𝐺𝛿 𝑓 = σ∞
𝑛=−∞ 𝑔 exp
2𝑊 𝑊
• Also, by 2), we get:
1
5) 𝐺 𝑓 = 2𝑊 𝐺𝛿 𝑓 , for −𝑊 < 𝑓 < 𝑊
• Substituting 4) into 5), we get:
1 𝑛 −𝑗𝜋𝑛𝑓
6) 𝐺 𝑓 = 2𝑊 σ∞
𝑛=−∞ 𝑔 exp , for −𝑊 < 𝑓 < 𝑊
2𝑊 𝑊
𝑛
• 6) shows that if sample values 𝑔 of signal 𝑔(𝑡) are specified for all 𝑛, then signal 𝑔(𝑡) is completely
2𝑊
determined for all values of 𝑡
• Taking inverse Fourier transform of 6), we get:
𝑛
7) 𝑔 𝑡 = σ∞
𝑛=−∞ 𝑔 2𝑊
sinc(2𝑊𝑡 − 𝑛) for 𝑡 ∈ (−∞, ∞)
• Equation 7) provides an interpolation formula for reconstructing the original signal 𝑔 𝑡 from the sequence
𝑛
of sample values 𝑔
2𝑊
• Thus, we have derived the “Sampling Theorem”, which states the following:
❑ A band-limited signal which only has frequency components in the range −𝑊 < 𝑓 < 𝑊 is completely described by
specifying the values of the signal at instants of time separated by 1/2𝑊 seconds
❑ Such a signal can be completely recovered from a knowledge of its samples taken at the rate of 2𝑊 samples per second
• Sampling rate of 2𝑊 samples per second, for a signal bandwidth of 𝑊 Hz, called Nyquist rate; its reciprocal
1
called Nyquist interval
2𝑊
 Aliasing
• In above derivation of sampling theorem, we assumed that signal
𝑔(𝑡) is strictly band-limited
• However, in practice, an information-bearing signal is not strictly
band-limited
❑ so some undersampling occurs
• So sampling process produces some “aliasing” as shown in fig
• To combat the effects of aliasing in
practice:
❑ Prior to sampling, a low-pass filter
used to attenuate those high-
frequency components that are not
essential to information being
conveyed by signal
❑ Filtered signal is sampled at a rate
slightly higher than Nyquist rate

Ref: “Communication Systems” by Haykin and Moher, 5th ed

 Aliasing (contd.)
• What is the benefit of using a sampling rate that is slightly higher than (not equal
to) Nyquist rate?
❑ Eases the design of the reconstruction filter used to recover original signal from its
sampled version
• E.g., suppose a message signal with bandwidth 𝑊 is sampled at rate 𝑓𝑠 > 2𝑊
• Then reconstruction filter:
❑ can be low-pass filter with a passband
extending from −𝑊 to 𝑊 and
❑ transition band extending (for positive
frequencies) from 𝑊 to 𝑓𝑠 − 𝑊 (see
fig)
• Thus, reconstruction filter allowed to
have transition band of width 𝑓𝑠 −
2𝑊 > 0
❑ In contrast, if 𝑓𝑠 = 2𝑊, then ideal
reconstruction filter with zero width
of transition band would be required,
which is not practically realizable

Ref: “Communication Systems” by Haykin and Moher, 5th ed

 Practical Sampling
• So far, we have considered ideal sampling
using an impulse pulse train
• But this sampling process is physically
unrealizable
• So next, we consider a practical
implementation of sampling
• Called “Pulse Amplitude Modulation”
 Pulse Amplitude Modulation (PAM)
• In PAM, amplitudes of regularly spaced pulses varied in proportion to
corresponding sample values of a continuous message signal 𝑚(𝑡) as
shown in fig.
❑ 𝑠(𝑡) is PAM signal obtained from 𝑚(𝑡)
• PAM signal 𝑠(𝑡) can be generated by following operations:
1) Instantaneous sampling of message signal 𝑚(𝑡) every 𝑇𝑠 seconds, where
sampling rate 𝑓𝑠 = 1/𝑇𝑠 chosen in accordance with sampling theorem
2) Lengthening duration of each sample to some constant value 𝑇
• Above two operations jointly referred to as “sample and hold”
• Reason for lengthening duration of each sample (step 2):
❑ To avoid use of excessive channel bandwidth
• PAM signal 𝑠(𝑡) can be expressed as:
❑ 𝑠 𝑡 = σ∞𝑛=−∞ 𝑚 𝑛𝑇𝑠 ℎ(𝑡 − 𝑛 𝑇𝑠 ),
1, 0 ≤ 𝑡 ≤ 𝑇,
❑ where ℎ 𝑡 = ቊ
0, else.
• Recall: 𝑚𝛿 𝑡 = σ∞𝑛=−∞ 𝑚 𝑛𝑇𝑠 𝛿(𝑡 − 𝑛 𝑇𝑠 )
• 𝑠 𝑡 in terms of 𝑚𝛿 𝑡 and ℎ 𝑡 :
❑ 𝑚𝛿 𝑡 ∗ ℎ(𝑡)
• Taking Fourier transforms on both sides:
• 𝑆 𝑓 = 𝑀𝛿 𝑓 𝐻(𝑓)
Ref: “Communication Systems” by Haykin and Moher, 5th ed
 Pulse Amplitude Modulation (PAM) (contd.)
• Recall:
❑ 𝑠 𝑡 = σ∞
𝑛=−∞ 𝑚 𝑛𝑇𝑠 ℎ(𝑡 − 𝑛 𝑇𝑠 ) = 𝑚𝛿 𝑡 ∗ ℎ(𝑡)
❑ 𝑆 𝑓 = 𝑀𝛿 𝑓 𝐻(𝑓)
❑ 𝑀𝛿 𝑓 = 𝑓𝑠 σ∞
𝑚=−∞ 𝑀 𝑓 − 𝑚𝑓𝑠
• So 𝑆 𝑓 :
❑ 𝑓𝑠 σ∞
𝑚=−∞ 𝑀 𝑓 − 𝑚𝑓𝑠 𝐻(𝑓)
• Given a PAM signal 𝑠 𝑡 , how can we recover
message signal 𝑚(𝑡)?
• Assuming that sampling rate exceeds Nyquist rate,
i.e., 𝑓𝑠 > 2𝑊, we pass 𝑠(𝑡) through low-pass filter to
get signal with Fourier transform 𝑀 𝑓 𝐻(𝑓)
1, 0 ≤ 𝑡 ≤ 𝑇,
• Recall: ℎ 𝑡 = ቊ
0, else.
• So 𝐻(𝑓):
❑ 𝑇sinc(𝑓𝑇)𝑒 −𝑗𝜋𝑓𝑇
• We can recover 𝑚(𝑡) by:
❑ passing the above signal with Fourier transform
𝑀 𝑓 𝐻(𝑓) through filter with amplitude response
1 1
=
|𝐻 𝑓 | |𝑇sinc 𝑓𝑇 |
• Fig. shows relevant amplitude spectra

Ref: “Communication Systems” by Haykin and Moher, 5th ed

 Communication Using Pulse Modulation
• Suppose a continuous-time message signal 𝑔(𝑡) needs to be sent over a baseband channel
• In “pulse modulation”:
❑ 𝑔(𝑡) is sampled
❑ sample values are used to modify certain parameters of a periodic pulse train
• Fig. shows:
❑ PAM signal, in which pulse amplitudes varied
❑ “Pulse Width Modulation (PWM)”, in which pulse widths varied
❑ “Pulse Position Modulation (PPM)”, in which pulse positions varied
• In all the above cases, instead of sending 𝑔(𝑡),
we transmit the corresponding pulse modulated
signal over channel
• Recall: previous slide shows that bandwidth of
PAM signal is larger than bandwidth of message
signal
• Advantage of pulse modulation over sending
message signal 𝑔(𝑡) itself:
❑ Pulse modulation allows simultaneous transmission
of several signals on a time-sharing basis, i.e., Time
Division Multiplexing (TDM), as shown in fig.

Ref: “Modern Digital and Analog Comm. Systems” by B.P. Lathi and Z. Ding, 4th ed
•
Quantization
Samples of a continuous signal, such as voice, take real values
❑ Hence, infinite number of amplitude levels
• Not necessary to transmit exact amplitudes of the samples
❑ Since human eye or ear, as final receiver, can only detect finite intensity differences
• So original continuous signal can be approximated by a signal constructed of
discrete amplitudes selected from a finite set as shown in fig
• If the discrete amplitude levels are assigned with sufficiently close spacing, then:
❑ approximated signal can be made practically indistinguishable from original continuous
signal
• “Quantization” defined as process of transforming the sample amplitude 𝑚 𝑛𝑇𝑠 of
a message signal 𝑚 𝑡 into a discrete amplitude 𝑣 𝑛𝑇𝑠 taken from a finite set of
possible amplitudes
• When digital communication used to
transmit an analog message source
(e.g., voice), then:
❑ After sampling, message signal is
quantized and then converted into a
sequence of bits

Ref: “Communication Systems” by Haykin and Moher, 5th ed

 Quantization (contd.)
• Recall: quantizer converts sample amplitude 𝑚 𝑛𝑇𝑠 of message signal
𝑚 𝑡 into amplitude 𝑣 𝑛𝑇𝑠 taken from a finite set of possible amplitudes
• For simplicity, drop time index and denote 𝑚 𝑛𝑇𝑠 and 𝑣 𝑛𝑇𝑠 by 𝑚 and 𝑣,
respectively
• For 𝑘 = 1, … , 𝐿: m, v are different from mk and vk
❑ if 𝑚 lies in the interval ℐ𝑘 = (𝑚𝑘 , 𝑚𝑘+1 ], then 𝑣 = 𝑣𝑘 as shown in fig
❑ where 𝐿 is total number of amplitude levels used in quantizer
❑ 𝑚𝑘 , 𝑘 = 1, … , 𝐿 + 1, called “decision thresholds”
❑ 𝑣𝑘 , 𝑘 = 1, … , 𝐿, called “representation levels”
• Spacing between adjacent representation levels called:
❑ “quantum” or “step-size”
• Quantizers can be of a uniform or
non-uniform type:
❑ in uniform quantizer,
representation levels are uniformly
spaced and in non-uniform
quantizer, they are not
• We first discuss uniform quantizers
and then non-uniform quantizers
Ref: “Communication Systems” by Haykin and Moher, 5th ed
Quantization (contd.)
• Quantizer characteristic can
be of following types (see
fig.):
❑“midtread type”: origin lies in
middle of a tread of the
staircase graph
❑“midrise type”: origin lies in
middle of a rising part of the
staircase graph

Ref: “Communication Systems” by Haykin and Moher, 5th ed

• Use of quantization introduces an error, given by: Quantization Noise
❑ 𝑞 = 𝑚 − 𝑣 m, v are different from mk, vk so this depends on message signal instead of defined
❑ called “quantization noise” values
• Suppose 𝑚, 𝑣, and 𝑞 are realizations of random variables 𝑀, 𝑉, and 𝑄, respectively, so that:
❑ 𝑄 =𝑀−𝑉 is this zero mean due to message signal being equally dist.
on either sides of zero or that the thresholds and
• For simplicity, assume that 𝑀 and 𝑉 are zero-mean representation levels are equally distributed on either sides
❑ so 𝑄 is also zero-mean of zero
• To find signal-to-quantization-noise ratio, we need to find 𝐸 𝑄 2 :
❑ i.e., variance of 𝑄, say 𝜎𝑄2
• Consider an input 𝑚 of continuous amplitude in range −𝑚𝑚𝑎𝑥 , 𝑚𝑚𝑎𝑥
• If 𝐿 is number of representation levels, then step-size:
2𝑚𝑚𝑎𝑥
❑ ∆=
𝐿
• In what range does 𝑄 lie?
∆ ∆
❑ −2,2
∆ ∆
• Assume that 𝑄 is uniformly distributed in − ,
2 2
• 𝜎𝑄2 :
∆2
❑ 12
• Typically, output of quantizer is transmitted to receiver in binary form
• Number of bits per sample, 𝑅:
❑ log 2 𝐿
• So 𝐿 = 2𝑅
• 𝜎𝑄2 in terms of 𝑚𝑚𝑎𝑥 and 𝑅:
2
1 𝑚𝑚𝑎𝑥
❑
3 22𝑅
• If average power of message signal 𝑚(𝑡) is 𝑃, then output SNR of quantizer:
3𝑃
❑ 2
𝑚𝑚𝑎𝑥
22𝑅
• Thus, output SNR of quantizer increases exponentially with increasing number of bits per sample, 𝑅
 Example
• Recall: if average power of message signal 𝑚(𝑡) is
𝑃, then output SNR of quantizer:
3𝑃
❑ 2 22𝑅
𝑚𝑚𝑎𝑥

• Consider sinusoidal message signal of amplitude

𝐴𝑚
• Output SNR:
3
❑ × 22𝑅
2
• Output SNR in dB:
❑1.761 + 6.02𝑅
 Pulse-Code Modulation (PCM)
• Message signal sampled and each sample is quantized and represented
using a set of bits
• Fig. shows transmitter of PCM system
• Low-pass filter used to:
❑ prevent aliasing
• Sampling (e.g., PAM) used with a sampling rate greater than Nyquist rate
• Quantization may be:
❑ uniform or
❑ non-uniform (discussed later)
• Encoding used to:
❑ convert the sequence of quantized values to a form that is suitable for
transmission over communication channel

Ref: “Communication Systems” by Haykin and Moher, 5th ed

 Non-Uniform Quantization
• Recall: in a uniform quantizer, there is a constant separation
between each pair of neighboring representation levels
• However, in certain applications, it is better to use variable
separations between the representation levels
• E.g.: range of voltages covered by voice signals, from peaks of loud
talk to weak passages of weak talk, is on the order of 1000 to 1
• For voice signal, non-uniform quantizer with the feature that step-
size increases as the separation from the origin of the input-output
characteristic increases is used
• Using above process, nearly uniform percentage precision is
achieved over the amplitude range of input signal
• Also, above process beneficial since for voice signal, there is higher
probability of signal having smaller magnitudes than higher
magnitudes
• Fewer steps are needed than in the case of:
❑ a uniform quantizer that achieves the same maximum percentage error
❑ this reduces the bandwidth requirement of quantized signal
 Compander
• Use of a non-uniform quantizer equivalent to:
❑ passing the input sample through a compressor and
❑ applying the compressed sample to a uniform quantizer
• Examples of compression laws that are widely used in practice:
❑ 𝜇-law
❑ 𝐴-law
• 𝜇-law defined by:
log 1+𝜇|𝑚|
1) 𝑣 = log 1+𝜇 ,
❑ where 𝑚 and 𝑣 are the normalized
input and output voltages of the
compressor
❑ and 𝜇 is a positive constant
• The case 𝜇 = 0 corresponds to:
❑ uniform quantization
• 𝜇-law in 1) is approximately linear at
low values of |𝑚| and logarithmic at
high values of |𝑚|
• 𝜇-law shown in fig Ref: “Communication Systems” by Haykin
• Practical values of 𝜇: around 100 and Moher, 5th ed
• 𝐴-law defined by: Compander (contd.)
𝐴|𝑚| 1
, 0≤ 𝑚 ≤ ,
1+log 𝐴 𝐴
1) 𝑣 = ൞1+log 𝐴|𝑚| 1
≤ |𝑚| ≤ 1.
1+log 𝐴 𝐴
• The case 𝐴 = 1 corresponds to:
❑ uniform quantization
• 𝐴-law shown in fig
• 𝐴-law in 1) is:
❑ linear for small values of 𝑚
❑ logarithmic for large values of 𝑚
• Practical values of 𝐴 : around 100
• When compressor used in transmitter:
❑ in receiver, we need a device to restore the
signal samples to their correct relative levels
❑ such a device called “expander”
• Compression and expansion laws are
approximately inverse of each other
• Combination of a compressor and expander
called a “compander” Ref: “Communication Systems” by Haykin and
Moher, 5th ed
 Encoding
• Encoding used to:
❑convert the sequence of quantized values to a form
that is suitable for transmission over communication
channel
• Typically, each quantized value converted to a
binary code:
❑e.g., if there are 16 representation levels, then first
level denoted by code 0000, second by 0001, third by
0010, sixteenth by 1111
• More generally, ternary codes or codes in which
more than three values are taken by each symbol
may also be used
• In a “line code”, a binary stream of data
Line Codes
takes on an electrical representation
• Fig. shows the waveforms of five common
line codes for the example data stream
01101001
• Line codes are of the nonreturn-to-zero
(NRZ) or return-to-zero (RZ) type
❑ RZ denotes that the pulse shape used to
represent the bit always returns to the 0
volts level by the end of the bit
❑ NRZ denotes that the pulse does not
necessarily return to the neutral level by
the end of the bit
a) Unipolar NRZ Signaling:
❑ Symbol 1 is represented by transmitting a
pulse of amplitude 𝐴 for the symbol
duration and symbol 0 by transmitting no
pulse
❑ Also referred to as “on-off signaling”
b) Polar NRZ Signaling:
❑ Symbols 1 and 0 represented by
transmitting pulses of amplitudes +𝐴 and
− 𝐴, respectively
Ref: “Communication Systems” by Haykin and Moher, 5th ed
 Line Codes (contd.)
c) Unipolar RZ Signaling:
❑ Symbol 1 represented by a rectangular
pulse of amplitude 𝐴 and half-symbol width
and symbol 0 by transmitting no pulse
• Polar RZ Signaling (not shown in fig.):
❑ symbol 1 (respectively, 0) represented by
rectangular pulse of amplutude 𝐴
(respectively, −𝐴) and half-symbol width
❑ Easy to extract clock signal at receiver for
timing synchronization:
o By taking absolute value of received signal
d) Bipolar RZ Signaling:
❑ Symbol 0 represented by transmitting no
pulse and symbol 1 alternately by
transmitting pulse of amplitude 𝐴 and half-
symbol width and pulse of amplitude −𝐴
and half-symbol width
• Bipolar RZ Signaling has the advantage over
Unipolar RZ signaling that:
❑ Former has 0 dc component and latter has
positive dc component
❑ DC component does not pass through some
components of a communication system
(e.g., transformer), which can cause
distortion of transmitted signal Ref: “Communication Systems” by Haykin and Moher, 5th ed
 Line Codes (contd.)
e) Split-Phase (Manchester
Code):
❑ symbol 1 represented by
positive pulse of amplitude
𝐴 followed by negative
pulse of amplitude −𝐴,
each of half symbol-width
❑ for symbol 0, polarities of
above two pulses reversed
• Desirable feature of
Manchester code:
❑Zero dc component
Ref: “Communication Systems” by Haykin and Moher, 5th ed
 Example: T1 System
• T1 carrier system designed to accommodate 24 voice channels for short-distance
communications
• Voice signal mostly limited to a band from 300 Hz to 3.1 kHz:
❑ so passed through a low-pass filter with cut-off frequency of 3.1 kHz before sampling
• Nyquist rate is 6.2 kHz
• Sampling rate of 8 kHz used
• For companding, T1 system uses a piecewise-linear characteristic (consisting of 15 linear
segments) to approximate the 𝜇-law with 𝜇 = 255
• There are a total of 255 possible amplitude levels
• So each quantized sample represented using an 8-bit codeword
• 24 voice channels combined using Time Division Multiplexing (TDM)
• Each frame of the multiplexed signal consists of:
❑ 24 8-bit words, one for each voice source,
❑ plus a single bit that is added to the end of the frame for synchronization (called “framing bit”)
• Hence, each frame consists of a total of 24 × 8 + 1 = 193 bits
• Each frame has a period of:
❑ 125 𝜇s, since sampling rate of 8 kHz used for each voice channel
• Hence, duration of each bit is 0.647 𝜇s
• Resultant transmission rate:
❑ 1.544 Mbps
• Framing bits allow receiver to find out where each frame begins
• Framing bits chosen such that a sequence of framing bits, one at the end of each frame,
forms a special pattern that is unlikely to be formed in a speech signal
❑ The special pattern is 100011011100
 Differential PCM
• PCM not a very efficient system; it generates a large number of bits and requires a lot of
bandwidth to transmit
• “Differential PCM”: a more bandwidth-efficient scheme than PCM
• Basic idea: given an analog source (e.g., speech signal):
❑ sample values are not independent
❑ instead, we can make a good guess about a sample value from past sample values
• Let 𝑚[𝑘] denote 𝑘’th sample of signal 𝑚(𝑡)
• Suppose instead of transmitting 𝑚[𝑘], we transmit the difference 𝑑 𝑘 = 𝑚 𝑘 − 𝑚[𝑘 − 1]
• At receiver, we can reconstruct 𝑚 𝑘
• Advantage of this scheme over transmitting 𝑚[𝑘]:
❑ differences between successive samples are generally much smaller than sample values
❑ so fewer bits are required to represent 𝑑 𝑘 than 𝑚 𝑘 , for a given signal-to-quantization-noise target
❑ equivalently, for a fixed number of bits, SNR improves
3𝑃
❑ e.g., recall that for uniform quantizer, output SNR of quantizer is 2 22𝑅 , which increases when
𝑚𝑚𝑎𝑥
𝑚𝑚𝑎𝑥 decreases
• Consider following improvement over above scheme: at transmitter, we estimate (predict)
the value of 𝑚[𝑘] from knowledge of previous several sample values
❑ let 𝑚[𝑘]
ෝ be predicted value
❑ we transmit 𝑑 𝑘 = 𝑚 𝑘 − 𝑚[𝑘]ෝ instead of 𝑚 𝑘
• If the prediction 𝑚[𝑘]
ෝ is good, then:
❑ 𝑚 𝑘 − 𝑚[𝑘]
ෝ will have smaller magnitude than 𝑚 𝑘 − 𝑚[𝑘 − 1]
❑ which will lead to further reduction in number of bits that need to be sent for a given SNR target
• This scheme, wherein 𝑑 𝑘 = 𝑚 𝑘 − 𝑚[𝑘]
ෝ is transmitted, is called “Differential PCM”
❑ first scheme described above is a special case of Differential PCM in which 𝑚
ෝ 𝑘 = 𝑚[𝑘 − 1]
• How can we predict the value of 𝑚[𝑘] from knowledge of previous several sample values?
• One way: using Taylor series:
❑ 𝑚 𝑡 + 𝑇𝑠 = 𝑚 𝑡 + 𝑇𝑠 𝑚ሶ 𝑡 +
𝑇𝑠2
𝑚ሷ 𝑡 +
𝑇𝑠3
𝑚
ഺ 𝑡 +⋯
Differential PCM
2 3!
• For small 𝑇𝑠 :
❑ 𝑚 𝑡 + 𝑇𝑠 ≈ 𝑚 𝑡 + 𝑇𝑠 𝑚ሶ 𝑡
(contd.)
• How can we approximately find 𝑚ሶ 𝑡 ?
1
❑ 𝑚ሶ 𝑘𝑇𝑠 ≈ 𝑚 𝑘𝑇𝑠 − 𝑚 (𝑘 − 1)𝑇𝑠
𝑇𝑠
• So: 𝑚[𝑘 + 1]:
1) ≈ 2𝑚 𝑘 − 𝑚[𝑘 − 1]
• 1) provides crude prediction for 𝑚[𝑘 + 1]
• We can improve prediction by using more terms from Taylor series
• So more generally, we can express the prediction formula as:
❑ 𝑚 𝑘 ≈ 𝑎1 𝑚 𝑘 − 1 + 𝑎2 𝑚 𝑘 − 2 + ⋯ + 𝑎𝑁 𝑚 𝑘 − 𝑁
• So our prediction is:
❑ 𝑚
ෝ 𝑘 = 𝑎1 𝑚 𝑘 − 1 + 𝑎2 𝑚 𝑘 − 2 + ⋯ + 𝑎𝑁 𝑚 𝑘 − 𝑁
• Transmitter quantizes and sends 𝑑 𝑘 = 𝑚 𝑘 − 𝑚[𝑘]
ෝ
• Difficulty with above scheme:
❑ Receiver does not know 𝑚 𝑘 − 1 , 𝑚 𝑘 − 2 , … , 𝑚 𝑘 − 𝑁 , but instead knows only their quantized
versions 𝑚𝑞 [𝑘 − 1], 𝑚𝑞 𝑘 − 2 , …, 𝑚𝑞 [𝑘 − 𝑁]
• So a better scheme is as follows
• Prediction at transmitter is:
❑ 𝑚
ෝ 𝑘 = 𝑎1 𝑚𝑞 𝑘 − 1 + 𝑎2 𝑚𝑞 𝑘 − 2 + ⋯ + 𝑎𝑁 𝑚𝑞 𝑘 − 𝑁
• If 𝑑 𝑘 = 𝑚 𝑘 − 𝑚[𝑘]
ෝ , then transmitter sends its quantized version, say 𝑑𝑞 [𝑘]
• Receiver finds 𝑚𝑞 𝑘 = 𝑑𝑞 𝑘 + 𝑚 ෝ 𝑘
• Note that in above scheme, the difference 𝑑 𝑘 is quantized, and not 𝑚 𝑘
• Amplitude of a signal 𝑚(𝑡) is in range −1 V to 1 V Example
• Max. frequency in the signal is 4 kHz
• It is transmitted using 8 bit/ sample PCM
• The same signal is transmitted using differential PCM, where the error signal 𝑑(𝑡)
ranges from −0.1 V to 0.1 V
• In both cases, uniform quantization is used
• Suppose step size in differential PCM case must be at most step size in PCM case
• Want to find transmission bit rate in each case and hence bit rate compression
ratio
• Bit rate in case of PCM:
❑ 64 kbps
• Step size in case of PCM:
2
❑ = 7.8125 mV
256
• If same step size used in differential PCM case, then number of quantization levels:
0.2
❑ = 25.6
7.8125×10−3
• Since number of quantization levels must be a power of 2, we use 5 bits/ sample in
differential PCM
• Bit rate in case of differential PCM:
❑ 40 kbps
• So bit rate compression ratio:
❑ 1.6
 Delta Modulation
• A message signal 𝑚(𝑡) that needs to be transmitted is oversampled, i.e., sampled
at a rate much higher than Nyquist rate
• Benefit of oversampling:
❑ increases the correlation between adjacent samples of the signal
❑ this is done to permit use of a simple quantizing strategy for constructing the encoded
signal
• Delta modulation (DM) provides a staircase approximation to oversampled version
of message signal as shown in Fig.
• Difference between input and the approximation is quantized into only two levels,
viz., ±∆:
❑ If approximation falls below signal at any sampling epoch, it is increased by ∆
❑ If approximation lies above signal, it is decreased by ∆
• If signal does not change
too rapidly from sample to
sample, staircase
approximation remains
within ±∆ of input signal

Ref: “Communication Systems” by Haykin

and Moher, 5th ed
 Delta Modulation (contd.)
• If 𝑚(𝑡) is input signal and 𝑚𝑞 (𝑡) is staircase approximation, then:
❑ 𝑒 𝑛𝑇𝑠 = 𝑚 𝑛𝑇𝑠 − 𝑚𝑞 (𝑛𝑇𝑠 − 𝑇𝑠 )
❑ 𝑒𝑞 𝑛𝑇𝑠 = ∆sgn 𝑒 𝑛𝑇𝑠
❑ 𝑚𝑞 𝑛𝑇𝑠 = 𝑚𝑞 𝑛𝑇𝑠 − 𝑇𝑠 + 𝑒𝑞 𝑛𝑇𝑠

• Fig. shows transmitter and

receiver of delta
modulation system
• Function of low-pass filter
in receiver:
❑ It has bandwidth equal to
that of 𝑚(𝑡)
❑ Removes noise in the
quantized staircase
waveform

Ref: “Communication Systems” by Haykin

and Moher, 5th ed
• Delta modulation is subject to two types of quantization error: Delta Modulation
1) Slope Overload Distortion
2) Granular Noise (contd.)
• 𝑚𝑞 𝑡 can change at the maximum rate:
∆
❑
𝑇𝑠
∆
• So if signal 𝑚(𝑡) varies too rapidly in comparison with the value 𝑇 , slope overload
𝑠
distortion (noise) occurs as shown in fig
• Sufficient condition to prevent slope overload distortion:
∆ 𝑑𝑚(𝑡)
❑ ≥ max
𝑇𝑠 𝑑𝑡
• Granular noise occurs when (see fig):
❑ Step size ∆ is too large relative to local slope characteristics of 𝑚(𝑡)
❑ This causes the staircase approximation 𝑚𝑞 𝑡 to hunt around a relatively flat segment of 𝑚(𝑡)
• Optimal step size ∆ needs to be chosen taking into account the trade-off between small
slope overload distortion and small granular noise
• Adaptive delta modulation:
❑ improved version of delta modulation in which step size ∆ varies with time in accordance with 𝑚(𝑡)

Ref: “Communication Systems” by Haykin

and Moher, 5th ed
 Example
• Delta modulation has smaller bandwidth requirement
than PCM as illustrated in this example
• For telephone applications, a typical PCM system uses
8 kHz sampling rate and an 8 bit representation of
each sample
• So bit rate:
❑64 kbps
• Sampling rates for delta modulation range from 16 to
32 kHz depending on the desired voice quality
• Corresponding bit rates:
❑16 to 32 kbps, since delta modulation requires 1 bit per
sample
• Bandwidth savings under delta modulation compared
with PCM:
❑50 to 75%