Visit : www.EasyEngineeering.

net

EC6502 DIGITAL SIGNAL PROCESSING

UNIT I DISCRETE FOURIER TRANSFORM 9
DFT and its properties, Relation between DTFT and DFT, FFT computations using
Decimation in time and Decimation in frequency algorithms, Overlap-add and save
methods

UNIT II INFINITE IMPULSE RESPONSE DIGITAL FILTERS 9

Review of design of analogue Butterworth and Chebyshev Filters, Frequency transformation in
analogue domain – Design of IIR digital filters using impulse invariance technique – Design of
digital filters using bilinear transform – pre warping – Realization using direct, cascade and
parallel forms.

UNIT III FINITE IMPULSE RESPONSE DIGITAL FILTERS 9

Symmetric and Antisymmetric FIR filters – Linear phase FIR filters – Design using
Hamming, Hanning and Blackmann Windows – Frequency sampling method – Realization of
FIR filters – Transversal, Linear phase and Polyphase structures.

UNIT IV FINITE WORD LENGTH EFFECTS 9

Fixed point and floating point number representations – Comparison – Truncation and
Rounding errors - Quantization noise – derivation for quantization noise power – coefficient
quantization error – Product quantization error - Overflow error – Roundoff noise power -
limit cycle oscillations due to product roundoff and overflow errors - signal scaling

UNIT V MULTIRATE SIGNAL PROCESSING 9

Introduction to Multirate signal processing-Decimation-Interpolation-Polyphase
implementation of FIR filters for interpolator and decimator -Multistage implementation of
sampling rate conversion- Design of narrow band filters - Applications of Multirate signal
processing.
TOTAL= 60 PERIODS

TEXT BOOKS
1. John G Proakis and Manolakis, “ Digital Signal Processing Principles, Algorithms and
Applications”, Pearson, Fourth Edition, 2007.
2. S.Salivahanan, A. Vallavaraj, C. Gnanapriya, Digital Signal Processing,
TMH/McGraw Hill International, 2007

REFERENCES
1. E.C.Ifeachor and B.W.Jervis,“Digital signal processing–A practical approach ”
Second edition, Pearson, 2002.
2. S.K. Mitra, Digital Signal Processing, A Computer Based approach, Tata
Mc GrawHill, 1998.
3. P.P.Vaidyanathan, Multirate Systems & Filter Banks, Prentice Hall, Englewood
cliffs, NJ, 1993.
4. Johny R. Johnson, Introduction to Digital Signal Processing, PHI, 2006.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

S.No Topic Page No.

UNIT I - FREQUENCY TRANSFORMATIONS 1
1.1 INTRODUCTION 1
1.2 DIFFERENCE BETWEEN FT & DFT 2
1.3 CALCULATION OF DFT & IDFT 2
1.4 DIFFERENCE BETWEEN DFT & IDFT 3
1.5 PROPERTIES OF DFT 4
1.6 APPLICATION OF DFT 10
1.7 FAST FOURIER ALGORITHM (FFT) 13
1.8 RADIX-2 FFT ALGORITHMS 14
1.9 COMPUTATIONAL COMPLEXITY FFT V/S DIRECT 18
COMPUTATION
1.10 BIT REVERSAL 19
1.11 DECIMATION IN FREQUENCY (DIFFFT) 20
1.12 GOERTZEL ALGORITHM 26
UNIT II - IIR FILTER DESIGN 29
2.1 INTRODUCTION 29
2.2 TYPES OF DIGITAL FILTER 33
2.3 STRUCTURES FOR FIR SYSTEMS 34
2.4 STRUCTURES FOR IIR SYSTEMS 36
2.5 CONVERSION OF ANALOG FILTER INTO DIGITAL FILTER 43
2.6 IIR FILTER DESIGN - BILINEAR TRANSFORMATION 44
METHOD (BZT)
2.7 LPF AND HPF ANALOG BUTTERWORTH FILTER TRANSFER 47
FUNCTION
2.8 METHOD FOR DESIGNING DIGITAL FILTERS USING BZT 47
2.9 BUTTERWORTH FILTER APPROXIMATION 48
2.10 FREQUENCY RESPONSE CHARACTERISTIC 48
2.11 FREQUENCY TRANSFORMATION 48
2.11.1 FREQUENCY TRANSFORMATION (ANALOG FILTER) 48
2.11.2 FREQUENCY TRANSFORMATION (DIGITAL FILTER) 48
UNIT III - FIR FILTER DESIGN 50
3.1 FEATURES OF FIR FILTER 50
3.2 SYMMETRIC AND ANTI-SYMMETRIC FIR FILTERS 50
3.3 GIBBS PHENOMENON 51
3.4 DESIGNING FILTER FROM POLE ZERO PLACEMENT 54
3.5 NOTCH AND COMB FILTERS 54
3.6 DIGITAL RESONATOR 55
UNIT IV – FINITE WORD LENGTH EFFECTS 80
4.1 NUMBER REPRESENTATION 80
4.1.1 FIXED-POINT QUANTIZATION ERRORS 80
4.1.2 FLOATING-POINT QUANTIZATION ERRORS 82
4.2 ROUNDOFF NOISE 83
4.3 ROUNDOFF NOISE IN FIR FILTERS 83
4.4 ROUNDOFF NOISE IN FIXED POINT FIR FILTERS 85

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

4.5 LIMIT CYCLE OSCILLATIONS 89

4.6 OVERFLOW OSCILLATIONS 90
4.7 COEFFICIENT QUANTIZATION ERROR 91
4.8 REALIZATION CONSIDERATIONS 93
UNIT V - APPLICATIONS OF DSP 95
5.1 SPEECH RECOGNITION 95
5.2 LINEAR PREDICTION OF SPEECH SYNTHESIS 95
5.3 SOUND PROCESSING 97
5.4 ECHO CANCELLATION 97
5.5 VIBRATION ANALYSIS 98
5.6 MULTISTAGE IMPLEMENTATION OF DIGITAL FILTERS 100
5.7 SPEECH SIGNAL PROCESSING 100
5.8 SUBBAND CODING 102
5.9 ADAPTIVE FILTER 107
5.10 AUDIO PROCESSING 110
5.10.1 MUSIC SOUND PROCESSING 110
5.10.2 SPEECH GENERATION 110
5.10.3 SPEECH RECOGNITION 111
5.11 IMAGE ENHANCEMENT 111

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

UNIT I
FREQUENCY TRANSFORMATIONS

PRE REQUISITE DISCUSSION:

The Discrete Fourier transform is performed for Finite length sequence whereas DTFT is used to
perform transformation on both finite and also Infinite length sequence. The DFT and the DTFT can be
viewed as the logical result of applying the standard continuous Fourier transform to discrete data. From
that perspective, we have the satisfying result that it's not the transform that varies.

1.1 INTRODUCTION

Any signal can be decomposed in terms of sinusoidal (or complex exponential)

components. Thus the analysis of signals can be done by transforming time domain signals into
frequency domain and vice-versa. This transformation between time and frequency domain is performed
with the help of Fourier Transform(FT) But still it is not convenient for computation by DSP processors
hence Discrete Fourier Transform(DFT) is used.

Time domain analysis provides some information like amplitude at sampling instant but does not
convey frequency content & power, energy spectrum hence frequency domain analysis is used.

For Discrete time signals x(n) , Fourier Transform is denoted as x(ω) & given by
∞
X(ω) = ∑ x (n) e –jωn FT….……(1)
n=-∞
DFT is denoted by x(k) and given by (ω= 2 ∏ k/N)
N-1
X(k) = ∑ x (n) e –j2 ∏ kn / N DFT…….(2)
n=0
IDFT is given as
N-1

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

x(n) =1/N ∑ X (k) e j2 ∏ kn / N IDFT……(3)

k=0

1.2 DIFFERENCE BETWEEN FT & DFT

Sr Fourier Transform (FT) Discrete Fourier Transform (DFT)

No
1 FT x(ω) is the continuous DFT x(k) is calculated only at discrete values
function of x(n). of ω. Thus DFT is discrete in nature.

2 The range of ω is from - ∏ to ∏ Sampling is done at N equally spaced points over

or 0 to 2∏. period 0 to 2∏. Thus DFT is sampled
version of FT.
3 FT is given by equation (1) DFT is given by equation (2)
4 FT equations are applicable to most DFT equations are applicable to causal, finite
of infinite sequences. duration sequences
5 In DSP processors & computers In DSP processors and computers DFT s are
applications of FT are limited mostly used.
because x(ω) is continuous APPLICATION
function of ω. a) Spectrum Analysis b)
Filter Design
Q) Prove that FT x(ω) is periodic with period 2∏.
Q) Determine FT of x(n)= an u(n) for -1< a < 1.
Q) Determine FT of x(n)= A for 0 ≤ n ≤ L-1.
Q) Determine FT of x(n)= u(n)
Q) Determine FT of x(n)= δ(n)
Q) Determine FT of x(n)= e–at u(t)

1.3 CALCULATION OF DFT & IDFT

For calculation of DFT & IDFT two different methods can be used. First method is using mathematical
equation & second method is 4 or 8 point DFT. If x(n) is the sequence of N samples then consider WN= e –
j2 ∏ / N (twiddle factor)

Four POINT DFT ( 4-DFT)

Sr No WN=W4=e –j ∏/2 Angle Real Imaginary Total

1 W4 0 0 1 0 1
2 W4 1 - ∏/2 0 -j -j
3 W4 2 -∏ -1 0 -1
4 W4 3 - 3 ∏ /2 0 J J

n=0 n=1 n=2 n=3

k=0 W4 0 W4 0 W4 0 W4 0
[WN] = k=1 W4 0 W4 1 W4 2 W4 3 k=2

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

W4 0 W4 2 W4 4 W4 6

k=3 W4 0 W4 3 W4 6 W4 9

Thus 4 point DFT is given as XN = [WN ] XN

1 1 1 1
[WN] = 1 –j -1 j
1 -1 1 -1
1 j -1 -j

EIGHT POINT DFT ( 8-DFT)

Sr No WN = W8= e –j Angle Magnitude Imaginary Total

∏/4
1 W8 0 0 1 ---- 1
2 W8 1 - ∏/4 1/√2 -j 1/√2 1/√2 -j 1/√2
3 W8 2 - ∏/2 0 -j -j
4 W8 3 - 3 ∏ /4 -1/√2 -j 1/√2 - 1/√2 -j 1/√2
5 W8 4 -∏ -1 ---- -1
6 W8 5 - 5∏ / 4 -1/√2 +j 1/√2 - 1/√2 + j
1/√2
7 W8 6 - 7∏ / 4 0 J J
8 W8 7 - 2∏ 1/√2 +j 1/√2 1/√2 + j 1/√2
Remember that W8 0 = W8 8 = W8 16 = W8 24 = W8 32 = W8 40 (Periodic Property)
Magnitude and phase of x(k) can be obtained as,
|x(k)| = sqrt ( Xr(k)2 + XI(k)2) Angle
x(k) = tan -1 (XI(k) / XR(k))

Examples:
Q) Compute DFT of x(n) = {0,1,2,3} Ans: x4=[6, -2+2j, -2, -2-2j ]
Q) Compute DFT of x(n) = {1,0,0,1} Ans: x4=[2, 1+j, 0, 1-j ]
Q) Compute DFT of x(n) = {1,0,1,0} Ans: x4=[2, 0, 2, 0 ]
Q) Compute IDFT of x(k) = {2, 1+j, 0, 1-j } Ans: x4=[1,0,0,1]

1.4 DIFFERENCE BETWEEN DFT & IDFT

Sr DFT (Analysis transform) IDFT (Synthesis transform)

No
1 DFT is finite duration discrete IDFT is inverse DFT which is used to calculate time
frequency sequence that is obtained domain representation (Discrete time
by sampling one period of FT. sequence) form of x(k).
2 DFT equations are applicable to causal finite IDFT is used basically to determine sample
duration sequences. response of a filter for which we know only
transfer function.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3 Mathematical Equation to calculate Mathematical Equation to calculate IDFT is given

DFT is given by by
N-1 N-1
X(k) = ∑ x (n) e –j2 ∏ kn / N x(n) = 1/N ∑ X (k)e j2 ∏ kn / N
n=0 n=0

4 Thus DFT is given by In DFT and IDFT difference is of factor 1/N &
X(k)= [WN][xn] sign of exponent of twiddle factor.
Thus
x(n)= 1/N [ WN]-1[XK]

1.5 PROPERTIES OF DFT

DFT
x(n) x(k)
N
1. Periodicity
Let x(n) and x(k) be the DFT pair then if
x(n+N) = x(n) for all n then
X(k+N) = X(k) for all k
Thus periodic sequence xp(n) can be given as
∞
xp(n) = ∑ x(n-lN)
l=-∞

2. Linearity
The linearity property states that if
DFT
x1(n) X1(k) And

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

N
DFT
x2(n) X2(k) Then
N
Then DFT
a1 x1(n) + a2 x2(n) a1 X1(k) + a2 X2(k) N
DFT of linear combination of two or more signals is equal to the same linear combination of DFT of
individual signals.

3. Circular Symmetries of a sequence

A) A sequence is said to be circularly even if it is symmetric about the point zero on the circle. Thus
X(N-n) = x(n)

B) A sequence is said to be circularly odd if it is anti symmetric about the point zero on the circle.
Thus X(N-n) = - x(n)

C) A circularly folded sequence is represented as x((-n))N and given by x((-n))N = x(N-n).

D) Anticlockwise direction gives delayed sequence and clockwise direction gives advance sequence.
Thus delayed or advances sequence x`(n) is related to x(n) by the circular shift.

4. Symmetry Property of a sequence

A) Symmetry property for real valued x(n) i.e xI(n)=0

This property states that if x(n) is real then X(N-k) = X*(k)=X(-k)

B) Real and even sequence x(n) i.e xI(n)=0 & XI(K)=0

This property states that if the sequence is real and even x(n)= x(N-n) then DFT becomes
N-1
X(k) = ∑ x(n) cos (2∏kn/N)
n=0

C) Real and odd sequence x(n) i.e xI(n)=0 & XR(K)=0

This property states that if the sequence is real and odd x(n)=-x(N-n) then DFT becomes
N-1
X(k) = -j ∑ x(n) sin (2∏kn/N)
n=0

D) Pure Imaginary x(n) i.e xR(n)=0

This property states that if the sequence is purely imaginary x(n)=j XI(n) then DFT
becomes
N-1
XR(k) = ∑ xI(n) sin (2∏kn/N)
n=0

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

N-1
XI(k) = ∑ xI(n) cos (2∏kn/N)
n=0
5. Circular Convolution
The Circular Convolution property states that if
DFT
x1(n) X1(k) And
N

DFT
x2(n) X2(k) Then
N
DFT
Then x1(n) N x2(n) x1(k) x2(k) N
It means that circular convolution of x1(n) & x2(n) is equal to multiplication of their DFT s.
Thus circular convolution of two periodic discrete signal with period N is given by
N-1
y(m) = ∑ x1 (n) x2 (m-n)N ……….(4)
n=0

Multiplication of two sequences in time domain is called as Linear convolution while Multiplication of
two sequences in frequency domain is called as circular convolution. Results of both are totally different
but are related with each other.

There are two different methods are used to calculate circular convolution
1) Graphical representation form
2) Matrix approach

DIFFERENCE BETWEEN LINEAR CONVOLUTION & CIRCULAR CONVOLUTION

Sr No Linear Convolution Circular Convolution

1 In case of convolution two signal sequences input Multiplication of two DFT s is called as
signal x(n) and impulse response h(n) given by circular convolution.
the same system, output y(n) is
calculated
2 Multiplication of two sequences in time Multiplication of two sequences in
domain is called as Linear convolution frequency domain is called as circular
convolution.
3 Linear Convolution is given by the equation y(n) Circular Convolution is calculated as
= x(n) * h(n) & calculated as N-1
∞ y(m) = ∑ x1 (n) x2 (m-n)N
y(n) = ∑ x (k) h(n – k ) n=0

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

k= -∞
4 Linear Convolution of two signals returns N-1 Circular convolution returns same
elements where N is sum of elements in both number of elements that of two signals.
sequences.

Q) The two sequences x1(n)={2,1,2,1} & x2(n)={1,2,3,4}. Find out the sequence x3(m)
which is equal to circular convolution of two sequences. Ans: X3(m)={14,16,14,16}

Q) x1(n)={1,1,1,1,-1,-1,-1,-1} & x2(n)={0,1,2,3,4,3,2,1}. Find out the sequence x3(m)

which is equal to circular convolution of two sequences. Ans: X3(m)={-4,-8,-8,-4,4,8,8,4}

Q) Perform Linear Convolution of x(n)={1,2} & h(n)={2,1} using DFT & IDFT.

Q) Perform Linear Convolution of x(n)={1,2,2,1} & h(n)={1,2,3} using 8 Pt DFT & IDFT.

DIFFERENCE BETWEEN LINEAR CONVOLUTION & CIRCULAR CONVOLUTION

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

6. Multiplication
The Multiplication property states that if
DFT
X1(n) x1(k) And
N

DFT
X2(n) x2(k) Then
N

DFT
Then x1(n) x2(n) 1/N x1(k) N N x2(k)

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

It means that multiplication of two sequences in time domain results in circular

convolution of their DFT s in frequency domain.

7. Time reversal of a sequence

The Time reversal property states that if
DFT
X(n) x(k) And
N
DFT
Then x((-n))N = x(N-n) x((-k))N = x(N-k)
N
It means that the sequence is circularly folded its DFT is also circularly folded.

8. Circular Time shift

The Circular Time shift states that if
DFT
X(n) x(k) And
N
DFT
Then x((n-l))N x(k) e –j2 ∏ k l / N
N
Thus shifting the sequence circularly by „l samples is equivalent to multiplying its DFT by
e –j2 ∏ k l / N
9. Circular frequency shift
The Circular frequency shift states that if
DFT
X(n) x(k) And
N
DFT
Then x(n) e j2 ∏ l n / N x((n-l))N
N
Thus shifting the frequency components of DFT circularly is equivalent to multiplying its time
domain sequence by e –j2 ∏ k l / N

10. Complex conjugate property

The Complex conjugate property states that if
DFT
X(n) x(k) then
N
DFT
x*(n) x*((-k))N = x*(N-k) And
N
DFT
x*((-n))N = x*(N-k) x*(k)
N
11. Circular Correlation

The Complex correlation property states

DFT
rxy(l) Rxy(k)= x(k) Y*(k)
N
Here rxy(l) is circular cross correlation which is given as
Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

N-1
rxy(l) = ∑ x (n) y*((n – l ))N
n=0

This means multiplication of DFT of one sequence and conjugate DFT of another sequence is equivalent
to circular cross-correlation of these sequences in time domain.

12.Parseval’sTheorem

The Parseval s theorem states

N-1 N-1
∑ X(n) y*(n) = 1/N ∑ x (k) y*(k)
n=0 n=0
This equation give energy of finite duration sequence in terms of its frequency components.

1.6 APPLICATION OF DFT

1. DFT FOR LINEAR FILTERING

Consider that input sequence x(n) of Length L & impulse response of same system is h(n) having M
samples. Thus y(n) output of the system contains N samples where N=L+M-1. If DFT of y(n) also
contains N samples then only it uniquely represents y(n) in time domain. Multiplication of two DFT s is
equivalent to circular convolution of corresponding time domain sequences. But the length of x(n) &
h(n) is less than N. Hence these sequences are appended with zeros to make their length N called as
“Zero padding”. The N point circular convolution and linear convolution provide the same
sequence. Thus linear convolution can be obtained by circular convolution. Thus linear filtering is
provided by
DFT.

When the input data sequence is long then it requires large time to get the output sequence. Hence other
techniques are used to filter long data sequences. Instead of finding the output of complete input
sequence it is broken into small length sequences. The output due to these small length sequences are
computed fast. The outputs due to these small length sequences are fitted one after another to get the
final output response.

METHOD 1: OVERLAP SAVE METHOD OF LINEAR FILTERING

Step 1> In this method L samples of the current segment and M-1 samples of the
previous segment forms the input data block. Thus data block will be

X1(n) ={0,0,0,0,0,………………… ,x(0),x(1),…………….x(L-1)}

X2(n) ={x(L-M+1), …………….x(L-1),x(L),x(L+1),,,,,,,,,,,,,x(2L-1)} X3(n)
={x(2L-M+1), …………….x(2L-1),x(2L),x(2L+2),,,,,,,,,,,,,x(3L-1)}

Step2> Unit sample response h(n) contains M samples hence its length is made N by padding
zeros. Thus h(n) also contains N samples.

h(n)={ h(0), h(1), …………….h(M-1), 0,0,0,……………………(L-1 zeros)}

Step3> The N point DFT of h(n) is H(k) & DFT of mth data block be xm(K) then
corresponding DFT of output be Y`m(k)

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Y`m(k)= H(k) xm(K)

Step 4> The sequence ym(n) can be obtained by taking N point IDFT of Y`m(k). Initial
(M-1) samples in the corresponding data block must be discarded. The last L samples are
the correct output samples. Such blocks are fitted one after another to get the final output.

X(n) of Size N
Size L

X1(n)

M-1 Size L
Zeros
X2(n)
Size L

X3(n)

Y1(n)

Discard M-1 Points

Y2(n)

Y3(n)

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Discard M-1 Points

Discard M-1 Points

Y(n) of Size N

METHOD 2: OVERLAP ADD METHOD OF LINEAR FILTERING

Step 1> In this method L samples of the current segment and M-1 samples of the
previous segment forms the input data block. Thus data block will be

X1(n) ={x(0),x(1),…………….x(L-1),0,0,0,……….}
X2(n) ={x(L),x(L+1),x(2L-1),0,0,0,0}
X3(n) ={x(2L),x(2L+2),,,,,,,,,,,,,x(3L-1),0,0,0,0}

Step2> Unit sample response h(n) contains M samples hence its length is made N by padding
zeros. Thus h(n) also contains N samples.

h(n)={ h(0), h(1), …………….h(M-1), 0,0,0,……………………(L-1 zeros)}

Step3> The N point DFT of h(n) is H(k) & DFT of mth data block be xm(K) then
corresponding DFT of output be Y`m(k)

Y`m(k)= H(k) xm(K)

Step 4> The sequence ym(n) can be obtained by taking N point IDFT of Y`m(k). Initial
(M-1) samples are not discarded as there will be no aliasing. The last (M-1) samples of
current output block must be added to the first M-1 samples of next output block. Such blocks are
fitted one after another to get the final output.

X(n) of Size N
Size L
M-1
X1(n) Zeros
Size L
M-1
X2(n) Zeros
Size L
M-1
X3(n) Zeros

Y1(n)

47
Y2(n)

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

M-1
Points add
together

Y(n) of Size N

DIFFERENCE BETWEEN OVERLAP SAVE AND OVERLAP ADD METHOD

Sr OVERLAP SAVE METHOD OVERLAP ADD METHOD

No
1 In this method, L samples of the current In this method L samples from input sequence
segment and (M-1) samples of the and padding M-1 zeros forms data block of size
previous segment forms the input data block. N.
2 Initial M-1 samples of output sequence are
discarded which occurs due to aliasing effect. There will be no aliasing in output data blocks.
3 To avoid loss of data due to aliasing last
M-1 samples of each data record are saved. Last M-1 samples of current output
block must be added to the first M-1 samples of
next output block. Hence called as overlap add
method.

2. SPECTRUM ANALYSIS USING DFT

DFT of the signal is used for spectrum analysis. DFT can be computed on digital computer or digital
signal processor. The signal to be analyzed is passed through anti-aliasing filter and samples at the rate
of Fs≥ 2 Fmax. Hence highest frequency component is Fs/2.

Frequency spectrum can be plotted by taking N number of samples & L samples of waveforms. The
total frequency range 2∏ is divided into N points. Spectrum is better if we take large value of N & L But
this increases processing time. DFT can be computed quickly using FFT algorithm hence fast processing
can be done. Thus most accurate resolution can be obtained by increasing number of samples.

1.7 FAST FOURIER ALGORITHM (FFT)

1. Large number of the applications such as filtering, correlation analysis, spectrum analysis
require calculation of DFT. But direct computation of DFT require large number of computations and
hence processor remain busy. Hence special algorithms are developed to compute DFT quickly called
as Fast Fourier algorithms (FFT).

48

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

2. The radix-2 FFT algorithms are based on divide and conquer approach. In this method, the
N-point DFT is successively decomposed into smaller DFT s. Because of this decomposition, the
number of computations are reduced.

1.8 RADIX-2 FFT ALGORITHMS

1. DECIMATION IN TIME (DITFFT)

There are three properties of twiddle factor WN

1) W N = W NK (Periodicity Property)
k+N
2) W N =- (Symmetry Property)
k+N/2 WNK
2
3) WN = WN/2.

N point sequence x(n) be splitted into two N/2 point data sequences f1(n) and f2(n). f1(n) contains even
numbered samples of x(n) and f2(n) contains odd numbered samples of x(n). This splitted operation is
called decimation. Since it is done on time domain sequence it is called “Decimation in Time”.
Thus

f1(m)=x(2m) where n=0,1,………….N/2-1

f2(m)=x(2m+1) where n=0,1,………….N/2-1
N point DFT is given as

N-1
kn
X(k) =∑ x (n) WN
(1)
n=0
Since the sequence x(n) is splitted into even numbered and odd numbered samples, thus

N/2-1 N/2-1
2m
k k(2m+1)
X(k) =∑ x (2m) WN + ∑ x (2m+1) WN (2)
m=0 m=0
k
X(k) =F1(k) + WN F2(k) (3)
X(k+N/2) =F1(k) - WN k F2(k) (Symmetry property) (4)

Fig 1 shows that 8-point DFT can be computed directly and hence no reduction in
computation.

x(0) X(0)
x(1) X(1)
x(2) X(2)
8 Point
x(3) X(3)
DFT

x(7) X(7)

49

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Fig 1. DIRECT COMPUTATION FOR N=8

50

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

x(0) f1(0) X(0)

x(2) N/2 Point f1(1) X(1)
x(4) f1(2) X(2)
x(6) DFT f1(3) X(3)

x(1) f2(0) w8 0 X(4)

x(3) f2(1) w8 1 X(5)
x(5) N/2 Point f2(2) w8 2 X(6)
x(7) DFT f2(3) w8 3 X(7)

Fig 2. FIRST STAGE FOR FFT COMPUTATION FOR N=8

Fig 3 shows N/2 point DFT base separated in N/4 boxes. In such cases equations become
2
g1(k) =P1(k) + WN k P2(k) (5)
g1(k+N/2) =p1(k) - WN 2k P2(k) (6)

x(0) F(0)
N/4 Point
x(4) F(1)
DFT

x(2) w80 F(2)

N/4 Point
x(6) DFT w82 F(3)

Fig 3. SECOND STAGE FOR FFT COMPUTATION FOR N=8

a A= a + W Nr b

b WNr B= a - W Nr b

Fig 4. BUTTERFLY COMPUTATION (THIRD STAGE)

x(0) A X(0)

x(2) w80 B X(1)

x(1) C w80 X(2)

x(3) w80 D w81 X(3)

51

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Fig 5. SIGNAL FLOW GRAPH FOR RADIX- DIT FFT N=4

x(0) A1 A2 X(0)

x(4) w80 B1 B2 X(1)

x(2) C1 w80 C2 X(2)

x(6) w80 D1 w82 D2 X(3)

x(1) E1 E2 w80 X(4)

x(5) w80 F1 F2 w81 X(5)

x(3) G1 w80 G2 w82 X(6)

x(7) w80 H1 w82 H2 w83 X(7)

Fig 6. SIGNAL FLOW GRAPH FOR RADIX- DIT FFT N=8

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1.9 COMPUTATIONAL COMPLEXITY FFT V/S DIRECT COMPUTATION

For Radix-2 algorithm value of N is given as N= 2V

Hence value of v is calculated as
V= log10 N / log10 2
= log2 N
Thus if value of N is 8 then the value of v=3. Thus three stages of decimation. Total
number of butterflies will be Nv/2 = 12.
If value of N is 16 then the value of v=4. Thus four stages of decimation. Total number of butterflies
will be Nv/2 = 32.

Each butterfly operation takes two addition and one multiplication operations. Direct computation
requires N2 multiplication operation & N2 – N addition operations.

N Direct computation DIT FFT algorithm Improvement in

Complex Complex Complex Complex processing speed for
Multiplication Addition Multiplication Addition multiplication
N2 N2 - N N/2 log2 N N log2 N
8 64 52 12 24 5.3 times
16 256 240 32 64 8 times
256 65536 65280 1024 2048 64 times

MEMORY REQUIREMENTS AND IN PLACE COMPUTATION

a A= a + W NrN b

b WNr B= a - W r b

Fig. BUTTERFLY COMPUTATION

From values a and b new values A and B are computed. Once A and B are computed, there is no
need to store a and b. Thus same memory locations can be used to store A

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

and B where a and b were stored hence called as In place computation. The advantage of in place
computation is that it reduces memory requirement.

Thus for computation of one butterfly, four memory locations are required for storing two complex
numbers A and B. In every stage there are N/2 butterflies hence total 2N memory locations are
required. 2N locations are required for each stage. Since stages are computed successively these memory
locations can be shared. In every stage N/2 twiddle factors are required hence maximum storage
requirements of N point DFT will be (2N + N/2).

1.10 BIT REVERSAL

For 8 point DIT DFT input data sequence is written as x(0), x(4), x(2), x(6), x(1), x(5), x(3), x(7) and the
DFT sequence X(k) is in proper order as X(0), X(1), X(2), X(3), X(4), x(5), X(6), x(7). In DIF FFT it is
exactly opposite. This can be obtained by bit reversal method.

Decimal Memory Address x(n) in Memory Address in bit New

binary (Natural Order) reversed order Address in
decimal
0 0 0 0 0 0 0 0
1 0 0 1 1 0 0 4
2 0 1 0 0 1 0 2
3 0 1 1 1 1 0 6
4 1 0 0 0 0 1 1
5 1 0 1 1 0 1 5
6 1 1 0 0 1 1 3
7 1 1 1 1 1 1 7

Table shows first column of memory address in decimal and second column as binary. Third
column indicates bit reverse values. As FFT is to be implemented on digital computer simple integer
division by 2 method is used for implementing bit reversal algorithms. Flow chart for Bit reversal
algorithm is as follows

DECIMAL
NUMBER B TO BE
REVERSED

I=1
B1=B
BR=0

B2=Int(B1/2) BR=(2*BR)+ B1- (2 *

B2)]
B1=B2; 53
I++

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Is
I > log2N

Store BR as Bit
reversal of B

1.11. DECIMATION IN FREQUENCY (DIFFFT)

In DIF N Point DFT is splitted into N/2 points DFT s. X(k) is splitted with k even and k odd
this is called Decimation in frequency(DIF FFT).

N point DFT is given as

N-1
X(k) =∑ x (n) WN kn
(1)
n=0
Since the sequence x(n) is splitted N/2 point samples, thus

N/2-1 N/2-1
X(k) =∑ x (n) WN + ∑ x (n + N/2) WN (2)
m=0 m=0

N/2-1 N/2-1 N
X(k) =∑ x (n) W kn
N + W NkN/2 ∑ x (n + N/2) W kn
m=0 m=0

N/2-1 N/2-1 N
X(k) =∑ x (n) W kn
N + (-1)k ∑ x (n + N/2) W kn
m=0 m=0

N/2-1
X(k) =∑ x (n) + (-1)k x(n + N/2) kn (3)
WN
m=0

Let us split X(k) into even and odd numbered samples

N/2-1
X(2k) =∑ x (n) + (-1)2k x(n + N/2) WN 2kn (4)
m=0

54

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

N/2-1
X(2k+1) =∑ x (n)+(-1)(2k+1) x(n + N/2)WN (2k+1)n (5)
m=0

Equation (4) and (5) are thus simplified as

g1(n) = x (n) + x(n + N/2)

g2(n) = x (n) - x(n + N/2) WN n

Fig 1 shows Butterfly computation in DIF FFT.

a A= a + b

b
WNr B= (a –b)W rN

Fig 1. BUTTERFLY COMPUTATION

Fig 2 shows signal flow graph and stages for computation of radix-2 DIF FFT algorithm of
N=4

x(0) A X(0)

x(1) B w40 X(2)

x(2) w40 C X(1)

x(3) w41 D w40 X(3)

Fig 2. SIGNAL FLOW GRAPH FOR RADIX- DIF FFT N=4

55

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Fig 3 shows signal flow graph and stages for computation of radix-2 DIF FFT algorithm of
N=8

x(0) A1 A2 X(0)

x(1) B1 B2 w80 X(4)

x(2) C1 w80 C2 X(2)

56

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

x(3) D1 w82 D2 w80 X(6)

x(4) w80 E1 E2 X(1)

x(5) w81 F1 F2 w80 X(5)

x(6) w82 G1 w80 G2 X(3)

x(7) w83 H1 w82 H2 w80 X(7)

Fig 3. SIGNAL FLOW GRAPH FOR RADIX- DIF FFT N=8

DIFFERENCE BETWEEN DITFFT AND DIFFFT

Sr No DIT FFT DIF FFT

1 DITFFT algorithms are based upon DIFFFT algorithms are based upon
decomposition of the input sequence decomposition of the output sequence
into smaller and smaller sub into smaller and smaller sub sequences.
sequences.
2 In this input sequence x(n) is splitted into In this output sequence X(k) is
even and odd numbered samples considered to be splitted into even and
odd numbered samples
3 Splitting operation is done on time Splitting operation is done on frequency
domain sequence. domain sequence.
4 In DIT FFT input sequence is in bit In DIFFFT, input sequence is in natural
reversed order while the output order. And DFT should be read in bit
sequence is in natural order. reversed order.

DIFFERENCE BETWEEN DIRECT COMPUTATION & FFT

Sr No Direct Computation Radix -2 FFT Algorithms

1 Direct computation requires large Radix-2 FFT algorithms requires less
number of computations as compared number of computations.
with FFT algorithms.
2 Processing time is more and more for large Processing time is less hence these
number of N hence processor algorithms compute DFT very quickly as
remains busy. compared with direct computation.

57

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3 Direct computation does not requires Splitting operation is done on time

splitting operation. domain basis (DIT) or frequency domain
basis (DIF)
4 As the value of N in DFT increases, the As the value of N in DFT increases, the
efficiency of direct computation decreases. efficiency of FFT algorithms increases.

5 In those applications where DFT is to be Applications

computed only at selected values 1) Linear filtering
of k(frequencies) and when these 2) Digital filter design
values are less than log2N then direct
computation becomes more efficient than
FFT.

Q) x(n)={1,2,2,1} Find X(k) using

DITFFT. Q) x(n)={1,2,2,1} Find X(k)
using DIFFFT.
Q) x(n)={0.3535,0.3535,0.6464,1.0607,0.3535,-1.0607,-1.3535,-0.3535} Find X(k)
using DITFFT.
Q) Using radix 2 FFT algorithm, plot flow graph for N=8.

58

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

59

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1.12 GOERTZEL ALGORITHM:

FFT algorithms are used to compute N point DFT for N samples of the sequence x(n). This requires N/2
log2N number of complex multiplications and N log2N complex additions. In some applications DFT is
to be computed only at selected values of frequencies and selected values are less than log2N, then
direct computations of DFT becomes more efficient than FFT. This direct computations of DFT can
be realized through linear filtering of x(n). Such linear filtering for computation of DFT can be
implemented using Goertzel algorithm.

60

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

By definition N point DFT is given as

N-1
k
X(k) =∑ x (m) WN m (1)
m=0

Multiplying both sides by WN-kN (which is always equal to 1).

N-1
X(k) =∑ x (m) W k(N-m)
N (2)
m=0
Thus for LSI system which has input x(n) and having unit sample response
-
hk(n)= WNkn u(n)

X(n) hk(n)= W -kn

N u(n) yk(n)

Linear convolution is given by

∞
y(n) = ∑ x (k) h(n – k )
k=- ∞

∞
yk(n) = ∑ x (m) WN u(n–m) (3)
m=-∞

As x(m) is given for N values

N-1
yk(n) = ∑ x (m) WN - (4)
m=0

The output of LSI system at n=N is given by

∞
yk(n)|n=N = ∑ x (m) WN-k(N-m) (5)
m=-∞

Thus comparing equation (2) and (5), X(k)

= yk(n)| n=N

Thus DFT can be obtained as the output of LSI system at n=N. Such systems can give
X(k) at selected values of k. Thus DFT is computed as linear filtering operations by
Goertzel Algorithm.

GLOSSARY:
Fourier Transform:

The Transform that used to analyze the signals or systems Characteristics in frequency domain , which
is difficult in case of Time Domain.
61

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Laplace Transform:

Laplace Transform is the basic continuous Transform. Then it is developed to represent the continuous
signals in frequency domain.

Discrete Time Fourier Transform:

For analyzing the discrete signals, the DTFT (Discrete Time Fourier Transform) is used. The output, that
the frequency is continuous in DTFT. But the Transformed Value should be discrete. Since the Digital
Signal Processors cannot work with the continuous frequency signals. So the DFT is developed to
represent the discrete signals in discrete frequency domain.

Discrete Fourier Transform:

Discrete Fourier Transform is used for transforming a discrete time sequence of finite length “N” into a
discrete frequency sequence of the same finite length “N”.

Periodicity:

If a discrete time signal is periodic then its DFT is also periodic. i.e. if a signal or sequence is
repeated after N Number of samples, then it is called periodic signal.

Symmetry:

If a signal or sequence is repeated its waveform in a negative direction after “N/2” number of Samples,
then it is called symmetric sequence or signal.

Linearity:

A System which satisfies the superposition principle is said to be a linear system. The DFT have the
Linearity property. Since the DFT of the output is equal to the sum of the DFT’s of the Inputs.

Fast Fourier Transform:

Fast Fourier Transform is an algorithm that efficiently computes the discrete fourier transform of a
sequence x(n). The direct computation of the DFT requires 2N2 evaluations of trignometric functions.
4N2 real multiplications and 4N(N-1) real additions.

62

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

UNIT II
IIR FILTER DESIGN

PREREQUISITE DISCUSSION:

Basically a digital filter is a linear time –invariant discrete time system. The terms Finite Impulse
response (FIR) and Infinite Impulse Response (IIR) are used to distinguish filter types. The FIR filters are
of Non-Recursive type whereas the IIR Filters are of recursive type.

2.1 INTRODUCTION

To remove or to reduce strength of unwanted signal like noise and to improve the quality of
required signal filtering process is used. To use the channel full bandwidth we mix up two or more
signals on transmission side and on receiver side we would like to separate it out in efficient way.
Hence filters are used. Thus the digital filters are mostly used in

1. Removal of undesirable noise from the desired signals

2. Equalization of communication channels
3. Signal detection in radar, sonar and communication
4. Performing spectral analysis of signals.

Analog and digital filters

In signal processing, the function of a filter is to remove unwanted parts of the signal, such as
random noise, or to extract useful parts of the signal, such as the components lying within a certain
frequency range.

The following block diagram illustrates the basic idea.

There are two main kinds of filter, analog and digital. They are quite different in their physical
makeup and in how they work.

An analog filter uses analog electronic circuits made up from components such as resistors,
capacitors and op amps to produce the required filtering effect. Such filter circuits are widely used in
such applications as noise reduction, video signal enhancement, graphic equalizers in hi-fi systems, and
many other areas.

In analog filters the signal being filtered is an electrical voltage or current which is the direct
analogue of the physical quantity (e.g. a sound or video signal or transducer output) involved.

63

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

A digital filter uses a digital processor to perform numerical calculations on sampled values of the
signal. The processor may be a general-purpose computer such as a PC, or a specialized DSP (Digital
Signal Processor) chip.

The analog input signal must first be sampled and digitized using an ADC (analog to digital
converter). The resulting binary numbers, representing successive sampled values of the input signal, are
transferred to the processor, which carries out numerical calculations on them. These calculations
typically involve multiplying the input values by constants and adding the products together. If
necessary, the results of these
calculations, which now represent sampled values of the filtered signal, are output through a DAC
(digital to analog converter) to convert the signal back to analog form.

In a digital filter, the signal is represented by a sequence of numbers, rather than a voltage or
current.

The following diagram shows the basic setup of such a system.

BASIC BLOCK DIAGRAM OF DIGITAL FILTERS

Analog signal
Sampler Quantizer Digital
Xa (t) & Encoder Filter

Discrete time Digital

signal signal

1. Samplers are used for converting continuous time signal into a discrete time signal by taking
samples of the continuous time signal at discrete time instants.

2. The Quantizer are used for converting a discrete time continuous amplitude signal into a digital
signal by expressing each sample value as a finite number of digits.

64

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3. In the encoding operation, the quantization sample value is converted to the binary equivalent of
that quantization level.

4. The digital filters are the discrete time systems used for filtering of sequences.
These digital filters performs the frequency related operations such as low pass, high pass,
band pass and band reject etc. These digital Filters are designed with digital hardware and
software and are represented by difference equation.

DIFFERENCE BETWEEN ANALOG FILTER AND DIGITAL FILTER

Sr Analog Filter Digital Filter

No
1 Analog filters are used for filtering Digital filters are used for filtering digital
analog signals. sequences.
2 Analog filters are designed with Digital Filters are designed with digital hardware like
various components like resistor, FF, counters shift registers, ALU and
inductor and capacitor software s like C or assembly language.
3 Analog filters less accurate & Digital filters are less sensitive to the environmental
because of component tolerance of active changes, noise and disturbances.
components & more sensitive to Thus periodic calibration can be avoided. Also
environmental changes. they are extremely stable.
4 Less flexible These are most flexible as software programs &
control programs can be easily modified. Several input
signals can be filtered by one digital filter.
5 Filter representation is in terms of Digital filters are represented by the difference
system components. equation.
6 An analog filter can only be changed by A digital filter is programmable, i.e. its operation is
redesigning the filter circuit. determined by a program stored in the processor's
memory. This means the digital filter
can easily be changed without affecting the
circuitry (hardware).

FILTER TYPES AND IDEAL FILTER CHARACTERISTIC:

Filters are usually classified according to their frequency-domain characteristic as lowpass, highpass,
bandpass and bandstop filters.
1. Lowpass Filter
A lowpass filter is made up of a passband and a stopband, where the lower frequencies
Of the input signal are passed through while the higher frequencies are attenuated.
|H (ω)|

ω
-ωc ωc

65

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

2. Highpass Filter
A highpass filter is made up of a stopband and a passband where the lower frequencies of the input
signal are attenuated while the higher frequencies are passed.
|H(ω)|

ω
-ωc ωc
3. Bandpass Filter
A bandpass filter is made up of two stopbands and one passband so that the lower and higher
frequencies of the input signal are attenuated while the intervening
frequencies are passed.
|H(ω)|

ω
-ω2 -ω1 ω2 ω1

4. Bandstop Filter
A bandstop filter is made up of two passbands and one stopband so that the lower and
higher frequencies of the input signal are passed while the intervening frequencies are
attenuated. An idealized bandstop filter frequency response has the following shape.
|H(ω)|

5. Multipass Filter
A multipass filter begins with a stopband followed by more than one passband. By default, a
multipass filter in Digital Filter Designer consists of three passbands and
four stopbands. The frequencies of the input signal at the stopbands are attenuated
while those at the passbands are passed.

6. Multistop Filter

66

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

A multistop filter begins with a passband followed by more than one stopband. By default, a
multistop filter in Digital Filter Designer consists of three passbands and two stopbands.

7. All Pass Filter

An all pass filter is defined as a system that has a constant magnitude response for all
frequencies.
|H(ω)| = 1 for 0 ≤ ω < ∏
The simplest example of an all pass filter is a pure delay system with system function
H(z) = Z-k. This is a low pass filter that has a linear phase characteristic.
All Pass filters find application as phase equalizers. When placed in cascade with a system that
has an undesired phase response, a phase equalizers is designed to
compensate for the poor phase characteristic of the system and therefore to produce an overall linear
phase response.
IDEAL FILTER CHARACTERISTIC

1. Ideal filters have a constant gain (usually taken as unity gain) passband
characteristic and zero gain in their stop band.
2. Ideal filters have a linear phase characteristic within their passband.
3. Ideal filters also have constant magnitude characteristic.
4. Ideal filters are physically unrealizable.

2.2 TYPES OF DIGITAL FILTER

Digital filters are of two types. Finite Impulse Response Digital Filter & Infinite Impulse
Response Digital Filter

DIFFERENCE BETWEEN FIR FILTER AND IIR FILTER

Sr FIR Digital Filter IIR Digital Filter

No
1 FIR system has finite duration unit sample IIR system has infinite duration unit
response. i.e h(n) = 0 for n<0 and n ≥ M sample response. i. e h(n) = 0 for n<0
Thus the unit sample response exists for the Thus the unit sample response exists for
duration from 0 to M-1. the duration from 0 to ∞.
2 FIR systems are non recursive. Thus output of FIR IIR systems are recursive. Thus they use
filter depends upon present and past feedback. Thus output of IIR filter
inputs. depends upon present and past inputs as well as
past outputs
3 Difference equation of the LSI system for Difference equation of the LSI system for
FIR filters becomes IIR filters becomes
M N M
y(n)=∑ bk x(n–k) y(n)=-∑ ak y(n–k)+∑ bk x(n–k)
k=0 k=1 k=0

4 FIR systems has limited or finite memory IIR system requires infinite memory.
requirements.

67

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

5 FIR filters are always stable Stability cannot be always guaranteed.

6 FIR filters can have an exactly linear phase IIR filter is usually more efficient design in
response so that no phase distortion is terms of computation time and
introduced in the signal by the filter. memory requirements. IIR systems usually
requires less processing time and storage as
compared with FIR.
7 The effect of using finite word length to Analogue filters can be easily and readily
implement filter, noise and quantization transformed into equivalent IIR digital
errors are less severe in FIR than in IIR. filter. But same is not possible in FIR
because that have no analogue
counterpart.
8 All zero filters Poles as well as zeros are present.
9 FIR filters are generally used if no phase IIR filters are generally used if sharp cutoff
distortion is desired. and high throughput is required.
Example: Example:
System described by System described by
Y(n) = 0.5 x(n) + 0.5 x(n-1) is FIR filter. Y(n) = y(n-1) + x(n) is IIR filter.
h(n)={0.5,0.5} h(n)=an u(n) for n≥0

2. 3 STRUCTURES FOR FIR SYSTEMS

FIR Systems are represented in four different ways

1. Direct Form Structures
2. Cascade Form Structure
3. Frequency-Sampling Structures
4. Lattice structures.

1. DIRECT FORM STRUCTURE OF FIR SYSTEM

The convolution of h(n) and x(n) for FIR systems can be written as
M-1
y(n)=∑ h(k) x(n–k) (1)
k=0

The above equation can be expanded as,

Y(n)= h(0) x(n) + h(1) x(n-1) + h(2) x(n-2) + …………… + h(M-1) x(n-M+1) (2)

Implementation of direct form structure of FIR filter is based upon the above equation.

x(n) x(n-1) x(n-M+1)

Z-1 Z-1
h(0) h(1) h(M-1)

+ + + +
h(0)x(n) h(0)x(n)+
h(1)x(n) y(n)

FIG - DIRECT FORM REALIZATION OF FIR SYSTEM

68

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1) There are M-1 unit delay blocks. One unit delay block requires one memory location.
Hence direct form structure requires M-1 memory locations.

2) The multiplication of h(k) and x(n-k) is performed for 0 to M-1 terms. Hence M
multiplications and M-1 additions are required.

3) Direct form structure is often called as transversal or tapped delay line filter.

2. CASCADE FORM STRUCTURE OF FIR SYSTEM

In cascade form, stages are cascaded (connected) in series. The output of one system is input to
another. Thus total K number of stages are cascaded. The total system function
'H' is given by

H= H1(z) . H2(z)……………………. Hk(z) (1)

H= Y1(z)/X1(z). Y2(z)/X2(z). ……………Yk(z)/Xk(z) (2)

k
H(z)=π Hk(z) (3)
k=1

x(n)=x1(n) y1(n)=x2(n) y2(n)=x3(n) yk(n)=y(n)

H1(z) H2(z) Hk(z)

FIG- CASCADE FORM REALIZATION OF FIR SYSTEM

Each H1(z), H2(z)… etc is a second order section and it is realized by the direct form as shown in
below figure.

System function for FIR systems

M-1
H(z)=∑ bk z-k (1)
k=0

Expanding the above terms we have

H(z)= H1(z) . H2(z)……………………. Hk(z)

where HK(z) = bk0 + bk1 z-1 + bk2 z-2 (2)

Thus Direct form of second order system is shown as

x(n) x(n-1)
Z-1 Z-1
bk0 bk1 bk2

+ + y(n)

FIG - DIRECT FORM REALIZATION OF FIR SECOND ORDER SYSTEM

69

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

2. 4 STRUCTURES FOR IIR SYSTEMS

IIR Systems are represented in four different ways

1. Direct Form Structures Form I and Form II
2. Cascade Form Structure
3. Parallel Form Structure
4. Lattice and Lattice-Ladder structure.

DIRECT FORM STRUCTURE FOR IIR SYSTEMS

IIR systems can be described by a generalized equations as

N M
y(n)=-∑ ak y(n–k)+∑ bk x(n–k) (1)
k=1 k=0

Z transform is given as
M N
H(z) = ∑ bk z / 1+ ∑ ak z–k
–k (2)
K=0 k=1

M N
Here H1(z) = ∑ bk z–k And H2(z) = 1+ ∑ ak z–k
K=0 k=0

Overall IIR system can be realized as cascade of two function H1(z) and H2(z). Here
H1(z) represents zeros of H(z) and H2(z) represents all poles of H(z).

DIRECT FORM - I
b0
x(n) y(n)
+ +

Z-1 Z-1
b1 -a1
+ +

Z-1 Z-1

b2 -a2
+ +

bM-1 -aN-1
+ +
Z-1 Z-1
bM -aN

70

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

FIG - DIRECT FORM I REALIZATION OF IIR SYSTEM

1. Direct form I realization of H(z) can be obtained by cascading the realization of

H1(z) which is all zero system first and then H2(z) which is all pole system.

2. There are M+N-1 unit delay blocks. One unit delay block requires one memory location.
Hence direct form structure requires M+N-1 memory locations.

3. Direct Form I realization requires M+N+1 number of multiplications and M+N

number of additions and M+N+1 number of memory locations.

DIRECT FORM - II

1. Direct form realization of H(z) can be obtained by cascading the realization of

H1(z) which is all pole system and H2(z) which is all zero system.

2. Two delay elements of all pole and all zero system can be merged into single delay element.

3. Direct Form II structure has reduced memory requirement compared to Direct form I
structure. Hence it is called canonic form.

4. The direct form II requires same number of multiplications(M+N+1) and additions

(M+N) as that of direct form I.

X(n) b0
y(n)
+ +

Z-1
-a1 b1
+ +

Z-1

-a2 b2
+ +

-aN-1 bN-1
+ +
Z-1
-aN bN

FIG - DIRECT FORM II REALIZATION OF IIR SYSTEM

71

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

CASCADE FORM STRUCTURE FOR IIR SYSTEMS

In cascade form, stages are cascaded (connected) in series. The output of one system is input to
another. Thus total K number of stages are cascaded. The total system function
'H' is given by

H= H1(z) . H2(z)……………………. Hk(z) (1)

H= Y1(z)/X1(z). Y2(z)/X2(z). ……………Yk(z)/Xk(z) (2)

k
H(z)=π Hk(z) (3)
k=1

x(n)=x1(n) y1(n)=x2(n) y2(n)=x3(n) yk(n)=y(n)

H1(z) H2(z) Hk(z)

FIG - CASCADE FORM REALIZATION OF IIR SYSTEM

Each H1(z), H2(z)… etc is a second order section and it is realized by the direct form as shown in
below figure.

System function for IIR systems

M N
H(z) = ∑ bk z–k / 1+ ∑ ak z–k (1)
K=0 k=1

Expanding the above terms we have

H(z)= H1(z) . H2(z)……………………. Hk(z)

where HK(z) = bk0 + bk1 z-1 + bk2 z-2 / 1 + ak1 z-1 + ak2 z-2 (2)

Thus Direct form of second order IIR system is shown as

X(n) bk0
y(n)
+ +

Z-1
-ak1 bk1
+ +

Z-1

-ak2 bk2
+ +

72

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

FIG - DIRECT FORM REALIZATION OF IIR SECOND ORDER SYSTEM (CASCADE)

PARALLEL FORM STRUCTURE FOR IIR SYSTEMS

System function for IIR systems is given as

M N
H(z) = ∑ bk z / 1+ ∑ ak z–k
–k (1)
K=0 k=1

= b0 + b1 z-1 + b2 z-2 + ……..+ bM z-M / 1 + a1 z-1 + a2 z-2 +……+ aN z-N (2)

The above system function can be expanded in partial fraction as follows

H(z) = C + H1(z) + H2(z)…………………….+ Hk(z) (3)

Where C is constant and Hk(z) is given as

Hk(z) = bk0 + bk1 z-1 / 1 + ak1 z-1 + ak2 z-2 (4)

H1(z) +

H2(z) +

X(n) y(n)
k1(z) +

FIG - PARALLEL FORM REALIZATION OF IIR SYSTEM

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

IIR FILTER DESIGN

1. IMPULSE INVARIANCE
2. BILINEAR TRANSFORMATION
3. BUTTERWORTH APPROXIMATION

4. IIR FILTER DESIGN - IMPULSE INVARIANCE METHOD

Impulse Invariance Method is simplest method used for designing IIR Filters. Important
Features of this Method are

1. In impulse variance method, Analog filters are converted into digital filter just by replacing
unit sample response of the digital filter by the sampled version of impulse response of analog
filter. Sampled signal is obtained by putting t=nT hence
h(n) = ha(nT) n=0,1,2. ………….
where h(n) is the unit sample response of digital filter and T is sampling interval.
2. But the main disadvantage of this method is that it does not correspond to simple algebraic
mapping of S plane to the Z plane. Thus the mapping from analog
frequency to digital frequency is many to one. The segments
(2k-1)∏/T ≤ Ω ≤ (2k+1) ∏/T of jΩ axis are all mapped on the unit circle ∏≤ω≤∏.
This takes place because of sampling.

3. Frequency aliasing is second disadvantage in this method. Because of frequency aliasing, the
frequency response of the resulting digital filter will not be identical to the original analog
frequency response.

4. Because of these factors, its application is limited to design low frequency filters like LPF or
a limited class of band pass filters.

RELATIONSHIP BETWEEN Z PLANE AND S PLANE

Z is represented as rejω in polar form and relationship between Z plane and S plane is given as
Z=eST where s= σ + j Ω.
Z= eST (Relationship Between Z plane and S plane)
Z= e (σ + j Ω) T
= eσ T . ej Ω T
Comparing Z value with the polar form we have.

r= e σ T and ω = Ω T
Here we have three condition
1) If σ = 0 then r=1
2) If σ < 0 then 0 < r < 1
3) If σ > 0 then r> 1
Thus

73

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1) Left side of s-plane is mapped inside the unit circle.

2) Right side of s-plane is mapped outside the unit circle.
3) jΩ axis is in s-plane is mapped on the unit circle.

Im(z)
1 jΩ

Re(z) σ

74

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Im(z)
1 jΩ

75

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Re(z) σ

2.5 CONVERSION OF ANALOG FILTER INTO DIGITAL FILTER

Let the system function of analog filter is n

Ha(s)= Σ Ck / s-pk (1)
k=1

where pk are the poles of the analog filter and ck are the coefficients of partial fraction expansion. The impulse
response of the analog filter ha(t) is obtained by inverse Laplace transform and given as

n
ha(t) = Σ Ck epkt (2)
k=1

The unit sample response of the digital filter is obtained by uniform sampling of ha(t). h(n) = ha(nT)
n=0,1,2. ………….

n
h(n) =Σ Ck epknT (3)
k=1

System function of digital filter H(z) is obtained by Z transform of h(n).

N ∞
H(z) =Σ Ck Σ epkT z-1 n (4)
k=1 n=0

Using the standard relation and comparing equation (1) and (4) system function of digital filter is given as

1 1

76

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

s - pk 1- epkT z-1

STANDARD RELATIONS IN IIR DESIGN

Sr No Analog System Function Digital System function

1
1 1
s-a 1- e z-1
aT

2
s+a 1- e-aT (cos bT) z-1
(s+a)2 + b2 1-2e -aT (cos bT)z-1+ e-2aTz-2
3
b e-aT (sin bT) z-1
(s+a) + b2
2
1-2e-aT (cos bT)z-1+ e-2aTz-2

EXAMPLES - IMPULSE INVARIANCE METHOD

Sr No Analog System Function Digital System function

1
s + 0.1 1 - (e -0.1Tcos3T)z-1
(s+0.1)2 + 9 1-2e-0.1T (cos 3T)z-1+ e-0.2Tz-
2
2 1
(s+1) (s+2) 0.148 z
(for sampling frequency of 5 z2 - 1.48 z + 0.548
samples/sec)
3 10
(s+2) 10
(for sampling time is 1 - z-1
0.01 sec)

2.6 IIR FILTER DESIGN - BILINEAR TRANSFORMATION METHOD (BZT)

The method of filter design by impulse invariance suffers from aliasing. Hence in order to overcome
this drawback Bilinear transformation method is designed. In analogue domain frequency axis is an
infinitely long straight line while sampled data z plane it is unit circle radius. The bilinear
transformation is the method of squashing the infinite straight analog frequency axis so that it becomes
finite.

Important Features of Bilinear Transform Method are

1. Bilinear transformation method (BZT) is a mapping from analog S plane to digital Z plane.
This conversion maps analog poles to digital poles and analog zeros to digital zeros. Thus all
poles and zeros are mapped.

2. This transformation is basically based on a numerical integration techniques used to simulate

an integrator of analog filter.
77

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3. There is one to one correspondence between continuous time and discrete time frequency
points. Entire range in Ω is mapped only once into the range -∏≤ω≤∏.

4. Frequency relationship is non-linear. Frequency warping or frequency

compression is due to non-linearity. Frequency warping means amplitude response of digital
filter is expanded at the lower frequencies and compressed at the higher
frequencies in comparison of the analog filter.

5. But the main disadvantage of frequency warping is that it does change the shape of the desired
filter frequency response. In particular, it changes the shape of the transition bands.

CONVERSION OF ANALOG FILTER INTO DIGITAL FILTER

Z is represented as rejω in polar form and relationship between Z plane and S plane in
BZT method is given as

2 z-1
S=
T z+1

2 rejω - 1
S=
T rejω + 1

2 r (cos ω + j sin ω) -1
S=
T r (cos ω + j sin ω) +1
2 r2 - 2r j 2 r sin ω
S= + p11 2
T 1+r2+2r cos ω +r +2r cos ω

Comparing the above equation with S= σ + j Ω. We have

2 r2 -1
σ=
T 1+ r2+2r cos ω

2 2 r sin ω
Ω=
T 1+ r2+2r cos ω

Here we have three condition

1) If σ < 0 then 0 < r < 1
2) If σ > 0 then r > 1
3) If σ = 0 then r=1

When r =1
Ω = 2 sin ω

78

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

T 1+cos ω

Ω=
= (2/T) tan (ω/2)

ω= 2 tan -1 (ΩT/2)

The above equations shows that in BZT frequency relationship is non-linear. The frequency
relationship is plotted as

ω 2 tan -1 (ΩT/2)

ΩT

FIG - MAPPING BETWEEN FREQUENCY VARIABLE ω AND Ω IN BZT METHOD.

DIFFERENCE - IMPULSE INVARIANCE Vs BILINEAR TRANSFORMATION

Sr Impulse Invariance Bilinear Transformation

No
1 In this method IIR filters are designed having a This method of IIR filters design is based on the
unit sample response h(n) that is trapezoidal formula for numerical
sampled version of the impulse response of integration.
the analog filter.
2 In this method small value of T is selected to The bilinear transformation is a conformal
minimize the effect of aliasing. mapping that transforms the j Ω axis into
the unit circle in the z plane only once, thus
avoiding aliasing of frequency components.

79

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3 They are generally used for low frequencies like For designing of LPF, HPF and almost all types of
design of IIR LPF and a limited class of Band pass and band stop filters this
bandpass filter method is used.
4 Frequency relationship is linear. Frequency relationship is non-linear.
Frequency warping or frequency
compression is due to non-linearity.
5 All poles are mapped from the s plane to the z All poles and zeros are mapped.
plane by the relationship
Zk= epkT. But the zeros in two domain does not
satisfy the same relationship.

2.7 LPF AND HPF ANALOG BUTTERWORTH FILTER TRANSFER FUNCTION

Sr Order of Low Pass Filter High Pass Filter

No the Filter
1 1 1 / s+1 s / s+1
2 2 1 / s2+ √2 s + 1 s2 / s2+ √2 s + 1

3 3 1 / s3 + 2 s2 + 2s +1 s3 / s3 + 2 s2 + 2s +1

2.8 METHOD FOR DESIGNING DIGITAL FILTERS USING BZT

step 1. Find out the value of ω *. c

*=
ωccc= (2/T) tan (ωc Ts/2)

step 2. Find out the value of frequency scaled analog transfer function

Normalized analog transfer function is frequency scaled by replacing s by s/ωp*.

step 3. Convert into digital filter

Apply BZT. i.e Replace s by the ((z-1)/(z+1)). And find out the desired transfer function of
digital function.

Example:
Q) Design first order high pass butterworth filter whose cutoff frequency is 1 kHz at sampling
frequency of 104 sps. Use BZT Method

Step 1. To find out the cutoff frequency

ωc = 2∏f
= 2000 rad/sec

Step 2. To find the prewarp frequency

ωc* = tan (ωc Ts/2)
= tan(∏/10)

Step 3. Scaling of the transfer function

For First order HPF transfer function H(s) = s/(s+1) Scaled

80

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

transfer function H*(s) = H(s) |s=s/ωc*

81

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

H*(s)= s/(s + 0.325)

Step 4. Find out the digital filter transfer function. Replace s by (z-1)/(z+1)

H(z)= z-1
1.325z -0.675

Q) Design second order low pass butterworth filter whose cutoff frequency is 1 kHz at sampling
frequency of 104 sps.

Q) First order low pass butterworth filter whose bandwidth is known to be 1 rad/sec . Use
BZT method to design digital filter of 20 Hz bandwidth at sampling frequency 60 sps.

Q) Second order low pass butterworth filter whose bandwidth is known to be 1 rad/sec . Use BZT
method to obtain transfer function H(z) of digital filter of 3 DB cutoff frequency of
150 Hz and sampling frequency 1.28 kHz.

Q) The transfer function is given as s2+1 / s2+s+1 The function is for Notch filter with frequency 1
rad/sec. Design digital Notch filter with the following specification
(1) Notch Frequency= 60 Hz
(2) Sampling frequency = 960 sps.

2.9 BUTTERWORTH FILTER APPROXIMATION

The filter passes all frequencies below Ωc. This is called passband of the filter. Also the filter
blocks all the frequencies above Ωc. This is called stopband of the filter. Ωc is called cutoff
frequency or critical frequency.
No Practical filters can provide the ideal characteristic. Hence approximation of the ideal
characteristic are used. Such approximations are standard and used for filter design.
Such three approximations are regularly used. a)
Butterworth Filter Approximation
b) Chebyshev Filter Approximation
c) Elliptic Filter Approximation
Butterworth filters are defined by the property that the magnitude response is maximally
flat in the passband.

|H(Ω)|2

ΩC Ω

82

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1
|Ha(Ω)|2=
1 + (Ω/Ωc)2N

The squared magnitude function for an analog butterworth filter is of the form.
1
|Ha(Ω)|2= 1 + (Ω/Ωc)2N

N indicates order of the filter and Ωc is the cutoff frequency (-3DB frequency).
At s = j Ω magnitude of H(s) and H(-s) is same hence

Ha(s) Ha(-s) = 1
1 + (-s2/Ω 2c)N

To find poles of H(s). H(-s) , find the roots of denominator in above equation.

-s2 = (-1)1/N
Ωc2

As e j(2k+1) ∏ = -1 where k = 0,1,2,…….. N-1.

-s2 = (e j(2k+1) )1/N

2 ∏
Ωc

s2 = (-1) Ω 2c e j(2k+1) ∏ / N

Taking the square root we get poles of s. p =

k

+ √-1 Ωc [ e j(2k+1) ∏ / N ]1/2

Pk = + j Ωc e j(2k+1) ∏ / 2N

As e j∏/2 = j

Pk = + Ωc e j∏/2 e j(2k+1) ∏ / 2N

Pk = + Ωc e j(N+2k+1) ∏ / 2N (1)
This equation gives the pole position of H(s) and H(-s).

2.10 FREQUENCY RESPONSE CHARACTERISTIC

The frequency response characteristic of |Ha(Ω)|2 is as shown. As the order of the filter N increases, the
butterworth filter characteristic is more close to the ideal characteristic. Thus at higher orders like N=16
the butterworth filter characteristic closely approximate ideal filter characteristic. Thus an infinite order
filter (N ∞) is required to get ideal characteristic.
83

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

|Ha(Ω)|2 N=18

84

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

N=6

N=2

|Ha(Ω)|

Ap

0.5

As
Ω
Ωp Ωc Ωs

Ap= attenuation in passband. As=

attenuation in stopband. Ωp =
passband edge frequency Ωs =
stopband edge frequency

Specification for the filter is

|Ha(Ω)| ≥ Ap for Ω ≤ Ωp and |Ha(Ω)| ≤ As for Ω ≥ Ωs. Hence we have
1
1 + (Ωp/Ωc)2N ≥ Ap2

1
1 + (Ωs/Ωc)2N ≤ As2

82

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

To determine the poles and order of analog filter consider equalities.

(Ωp/Ωc)2N = (1/Ap2) – 1 (Ωs/Ωc)2N = (1/As2) - 1

Ωs
= (1/As2)-1
2
N (1/Ap2)-1
Ωp

Hence order of the filter (N) is calculated as

log (1/As2)-1
(1/Ap2)-1 (2)
N= 0.5
log (Ωs/ Ωp)

N= 0.5 log((1/As2) -1) (2A)

(2) log (Ωs/ Ωc)

And cutoff frequency Ωc is calculated as

Ωp (3)
Ωc = [(1/Ap2) -1]1/2N

If As and Ap values are given in DB then

As (DB) = - 20 log As

log As = -As /20

As = 10 -As/20

(As)-2 = 10 As/10

(As)-2 = 10 0.1 As DB

Hence equation (2) is modified as

log 100.1 As -1
100.1 Ap -1 (4)
N= 0.5 log (Ωs/ Ωp)

83

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Q) Design a digital filter using a butterworth approximation by using impulse invariance.

Example

|Ha(Ω)|

0.89125

0.17783

0.2∏ 0.3∏ Ω
Filter Type - Low Pass Filter
Ap - 0.89125
As - 0.17783
Ωp - 0.2∏
Ωs - 0.3∏

Step 1) To convert specification to equivalent analog filter.

(In impulse invariance method frequency relationship is given as ω= Ω T while in Bilinear
transformation method frequency relationship is given as Ω= (2/T) tan (ω/2) If Ts is not
specified consider as 1)

|Ha(Ω)| ≥ 0.89125 for Ω ≤ 0.2∏/T and |Ha(Ω)| ≤ 0.17783 for Ω ≥ 0.3∏/T.

Step 2) To determine the order of the filter.

log (1/As2)-1
(1/Ap2)-1
N= 0.5 log (Ωs/ Ωp)

N= 5.88

A) Order of the filter should be integer.

B) Always go to nearest highest integer vale of N.

Hence N=6

Step 3) To find out the cutoff frequency (-3DB frequency)

Ωp
Ωc =
[(1/Ap2) -1]1/2N

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

cutoff frequency Ωc = 0.7032

Step 4) To find out the poles of analog filter system function.

Pk = + Ωc e j(N+2k+1) ∏ / 2N

As N=6 the value of k = 0,1,2,3,4,5.

K Poles
-0.182 + j 0.679
0 P0= + 0.7032 e j7∏/12 0.182 - j 0.679

-0.497 + j 0.497
1 P1= + 0.7032 e j9∏/12 0.497 - j 0.497

-0.679 + j 0.182
2 P2= + 0.7032 e j11∏/12 0.679 - j 0.182

3 P3= + 0.7032 e j13∏/12 -0.679 - j 0.182

0.679 + j 0.182
4
-0.497 - j 0.497
P4= + 0.7032 e j15∏/12 0.497 + j 0.497
5
-0.182 - j 0.679
P5= + 0.7032 e j17∏/12 0.182 + j 0.679

For stable filter all poles lying on the left side of s plane is selected. Hence

S1 = -0.182 + j 0.679 S1* = -0.182 - j 0.679

S2 = -0.497 + j 0.497 S2* = -0.497 - j 0.497
S3 = -0.679 + j 0.182 S3* = -0.679 - j 0.182

Step 5) To determine the system function (Analog Filter)

Ha(s) = Ωc6
(s-s1)(s-s1*) (s-s2)(s-s2*) (s-s3)(s-s3*)

Hence

Ha(s) = (0.7032)6
(s+0.182-j0.679)(s+0.182+j0.679) (s+0.497-j0.497) (s+0.497+j0.497)
(s+0.679-j0.182)(s+0.679-j0.182)

Ha(s) = 0.1209
[(s+0.182)2 +(0.679)2] [(s+0.497)2+(0.497)2] [(s+0.679)2-(0.182)2]

Ha(s) = 1.97 × 0.679 × 0.497 × 0.182

[(s+0.182) +(0.679) ] [(s+0.497)2+(0.497)2] [(s+0.679)2-(0.182)2]
2 2
Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Step 6) To determine the system function (Digital Filter)

(In Bilinear transformation replace s by the term ((z-1)/(z+1)) and find out the transfer function of
digital function)
Step 7) Represent system function in cascade form or parallel form if asked.

Q) Given for low pass butterworth filter

Ap= -1 db at 0.2∏
As= -15 db at 0.3∏
1) Calculate N and Pole location
2) Design digital filter using BZT method.

Q) Obtain transfer function of a lowpass digital filter meeting specifications

Cutoff 0-60Hz
Stopband > 85Hz
Stopband attenuation > 15 db
Sampling frequency= 256 Hz . use butterworth characteristic.

Q) Design second order low pass butterworth filter whose cutoff frequency is 1 kHz at sampling
frequency of 104 sps. Use BZT and Butterworth approximation.

2.11 FREQUENCY TRANSFORMATION

When the cutoff frequency Ωc of the low pass filter is equal to 1 then it is called normalized filter.
Frequency transformation techniques are used to generate High pass filter, Bandpass and bandstop
filter from the lowpass filter system function.

2.11.1 FREQUENCY TRANSFORMATION (ANALOG FILTER)

Sr No Type of transformation Transformation ( Replace s by)

s
1 Low Pass
ωlp
ωlp - Password edge frequency of another LPF

ωhp
2 High Pass
s
ωhp = Password edge frequency of HPF

(s2 + ωl ωh )
3 Band Pass s (ωh - ωl )
ωh - higher band edge frequency
ωl - Lower band edge frequency

s(ωh - ωl)
s2+ ωh ωl
4 Band Stop
ωh - higher band edge frequency
ωl - Lower band edge frequency
86

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

2.11.2 FREQUENCY TRANSFORMATION (DIGITAL FILTER)

Sr No Type of transformation Transformation ( Replace z-1 by)

1 Low Pass z-1 - a

1 - az-1

2 High Pass - (z-1+ a)

1 + az-1

- (z-2 - a1z-1 + a2)

3 Band Pass
a2z-2 - a1z-1 + 1

z-2 - a1z-1 + a2
4 Band Stop
a2z-2 - a1z-1 + 1

Example:
Q) Design high pass butterworth filter whose cutoff frequency is 30 Hz at sampling frequency
of 150 Hz. Use BZT and Frequency transformation.

Step 1. To find the prewarp cutoff frequency

ωc* = tan (ωcTs/2)
= 0.7265

Step 2. LPF to HPF transformation

For First order LPF transfer function H(s) = 1/(s+1) Scaled

transfer function H*(s) = H(s) |s=ωc*/s
H*(s)= s/(s + 0.7265)

Step 4. Find out the digital filter transfer function. Replace s by (z-1)/(z+1)
H(z)= z-1
1.7265z - 0.2735

Q) Design second order band pass butterworth filter whose passband of 200 Hz and 300
Hz and sampling frequency is 2000 Hz. Use BZT and Frequency transformation.

Q) Design second order band pass butterworth filter which meet following specification
Lower cutoff frequency = 210 Hz
Upper cutoff frequency = 330 Hz
Sampling Frequency = 960 sps
Use BZT and Frequency transformation.

87

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

GLOSSARY:

System Design:

Usually, in the IIR Filter design, Analog filter is designed, then it is transformed to a digital filter the
conversion of Analog to Digital filter involves mapping of desired digital filter specifications into equivalent
analog filter.

Warping Effect:

The analog Frequency is same as the digital frequency response. At high frequencies, the relation between
ω and Ω becomes Non-Linear. The Noise is introduced in the Digital Filter as in the Analog Filter. Amplitude
and Phase responses are affected by this warping effect.

Prewarping:

The Warping Effect is eliminated by prewarping of the analog filter. The analog frequencies are
prewarped and then applied to the transformation.

Infinite Impulse Response:

Infinite Impulse Response filters are a Type of Digital Filters which has infinite impulse response. This
type of Filters are designed from analog filters. The Analog filters are then transformed to Digital Domain.

Bilinear Transformation Method:

In Bilinear transformation method the transform of filters from Analog to Digital is carried out in a way
such that the Frequency transformation produces a Linear relationship between Analog and Digital Filters.

Filter:

A filter is one which passes the required band of signals and stops the other unwanted band of frequencies.

Pass band:

The Band of frequencies which is passed through the filter is termed as passband.

Stopband:

The band of frequencies which are stopped are termed as stop band.

88

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

89

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

UNIT III
FIR FILTER DESIGN
PREREQUISITE DISCUSSION:
The FIR Filters can be easily designed to have perfectly linear Phase. These filters can be realized
recursively and Non-recursively. There are greater flexibility to control the Shape of their Magnitude
response. Errors due to round off noise are less severe in FIR Filters, mainly because Feed back is not
used.

3.1 Features of FIR Filter

1. FIR filter always provides linear phase response. This specifies that the signals in the pass
band will suffer no dispersion Hence when the user wants no phase distortion, then FIR
filters are preferable over IIR. Phase distortion always degrade the system performance. In
various applications like speech processing, data transmission over long distance FIR filters
are more preferable due to this characteristic.

2. FIR filters are most stable as compared with IIR filters due to its non feedback nature.

3. Quantization Noise can be made negligible in FIR filters. Due to this sharp cutoff
FIR filters can be easily designed.

4. Disadvantage of FIR filters is that they need higher ordered for similar magnitude response of
IIR filters.

FIR SYSTEM ARE ALWAYS STABLE. Why?

Proof:
Difference equation of FIR filter of length M is given as
M-1
y(n)=∑ bk x(n–k) (1)
k=0

And the coefficient bk are related to unit sample response as

H(n) = bn for 0 ≤ n ≤ M-1
= 0 otherwise.

We can expand this equation as

Y(n)= b0 x(n) + b1 x(n-1) + …….. + bM-1 x(n-M+1) (2)

System is stable only if system produces bounded output for every bounded input. This is stability
definition for any system.
Here h(n)={b0, b1, b2, } of the FIR filter are stable. Thus y(n) is bounded if input x(n) is
bounded. This means FIR system produces bounded output for every bounded
input. Hence FIR systems are always stable.

3.2 Symmetric and Anti-symmetric FIR filters

1. Unit sample response of FIR filters is symmetric if it satisfies following condition.
h(n)= h(M-1-n) n=0,1,2…………….M-1

2. Unit sample response of FIR filters is Anti-symmetric if it satisfies following condition

h(n)= -h(M-1-n) n=0,1,2…………….M-1 90

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

FIR Filter Design Methods

The various method used for FIR Filer design are as follows
1. Fourier Series method
2. Windowing Method
3. DFT method
4. Frequency sampling Method. (IFT Method)

3.3 GIBBS PHENOMENON

Consider the ideal LPF frequency response as shown in Fig 1 with a normalizing angular cut off
frequency Ωc.

Impulse response of an ideal LPF is as shown in Fig 2.

1. In Fourier series method, limits of summation index is -∞ to ∞. But filter must have finite terms.
Hence limit of summation index change to -Q to Q where Q is some finite integer. But this type of
truncation may result in poor convergence of the series. Abrupt truncation of infinite series is equivalent
to multiplying infinite series with rectangular sequence. i.e at the point of discontinuity some oscillation
may be observed in resultant series.

2. Consider the example of LPF having desired frequency response Hd (ω) as shown
in figure. The oscillations or ringing takes place near band-edge of the filter.

3. This oscillation or ringing is generated because of side lobes in the frequency response
W(ω) of the window function. This oscillatory behavior is called "Gibbs Phenomenon".

91

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Truncated response and ringing effect is as shown in fig 3.

WINDO WING TECHNIQUE

W[n]
Windowing is the quickest method for designing an FIR filter. A windowing function simply
truncates the ideal impulse response to obtain a causal FIR approximation that is non causal and
infinitely long. Smoother window functions provide higher out-of band rejection in the filter response.
However this smoothness comes at the cost of wider stopband transitions.
Various windowing method attempts to minimize the width of the main lobe (peak) of the
frequency response. In addition, it attempts to minimize the side lobes (ripple) of the frequency
response.

Rectangular Window: Rectangular This is the most basic of windowing methods. It does not require
any operations because its values are either 1 or 0. It creates an abrupt discontinuity that results in sharp
roll-offs but large ripples.

Rectangular window is defined by the following equation.

=1 for 0 ≤ n ≤ N
=0 otherwise

Triangular Window: The computational simplicity of this window, a simple convolution of two
rectangle windows, and the lower sidelobes make it a viable alternative to the rectangular window.

92

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Kaiser Window: This windowing method is designed to generate a sharp central peak. It has reduced
side lobes and transition band is also narrow. Thus commonly used in FIR filter design.

Hamming Window: This windowing method generates a moderately sharp central peak. Its ability to
generate a maximally flat response makes it convenient for speech processing filtering.

Hanning Window: This windowing method generates a maximum flat filter design.

93

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

3.4 DESIGNING FILTER FROM POLE ZERO PLACEMENT

Filters can be designed from its pole zero plot. Following two constraints should be imposed
while designing the filters.

1. All poles should be placed inside the unit circle on order for the filter to be stable. However zeros
can be placed anywhere in the z plane. FIR filters are all zero filters hence they are always stable. IIR
filters are stable only when all poles of the filter are inside unit circle.

2. All complex poles and zeros occur in complex conjugate pairs in order for the filter coefficients
to be real.

In the design of low pass filters, the poles should be placed near the unit circle at points corresponding to
low frequencies ( near ω=0)and zeros should be placed near or on unit circle at points corresponding to
high frequencies (near ω=∏). The opposite is true for high pass filters.

3.5 NOTCH AND COMB FILTERS

A notch filter is a filter that contains one or more deep notches or ideally perfect nulls in its frequency
response characteristic. Notch filters are useful in many applications where specific
frequency components must be eliminated. Example Instrumentation and
recording systems required that the power-line frequency 60Hz and its harmonics be eliminated.
To create nulls in the frequency response of a filter at a frequency ω0, simply introduce a
pair of complex-conjugate zeros on the unit circle at an angle ω0.
comb filters are similar to notch filters in which the nulls occur periodically across the
frequency band similar with periodically spaced teeth. Frequency response
characteristic of notch filter |H(ω)| is as shown

94

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

ωo ω1 ω

3.6 DIGITAL RESONATOR

A digital resonator is a special two pole bandpass filter with a pair of complex conjugate poles
located near the unit circle. The name resonator refers to the fact that the filter has a larger magnitude
response in the vicinity of the pole locations. Digital resonators are useful in many applications,
including simple bandpass filtering and speech generations.

IDEAL FILTERS ARE NOT PHYSICALLY REALIZABLE. Why?

Ideal filters are not physically realizable because Ideal filters are anti-causal and as only causal
systems are physically realizable.

Proof:
Let take example of ideal lowpass filter.
H(ω) = 1 for - ωc ≤ ω ≤ ωc
= 0 elsewhere

The unit sample response of this ideal LPF can be obtained by taking IFT of H(ω).
∞
1_
h(n) = 2∏ ∫ H(ω) ejωn dω (1)
-∞

ωc
h(n) = 1_ ∫ 1 ejωn dω (2)
2∏ -ωc

ωc
1_
h(n) = 2∏ ejωn
jn
-ωc

1
[ejωcn - e-jωcn ]
2∏jn

Thus h(n)= sin ωcn / ∏n for n≠0

Putting n=0 in equation (2) we have
ωc
h(n) = 1_ ∫ 1 dω (3)
2∏ -ωc

95

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

1_ [ω] ωc
2∏ -ωc

and h(n) = ωc / ∏ for n=0

i.e
sin (ωcn )
∏n for n≠0

h(n) =
ωc
n for n=0

Hence impulse response of an ideal LPF is as shown in Fig

LSI system is causal if its unit sample response satisfies following condition. h(n) = 0
for n<0
In above figure h(n) extends -∞ to ∞. Hence h(n) ≠0 for n<0. This means causality condition is not
satisfied by the ideal low pass filter. Hence ideal low pass filter is non
causal and it is not physically realizable.

EXAMPLES OF SIMPLE DIGITAL FILTERS:

The following examples illustrate the essential features of digital filters.

1. UNITY GAIN FILTER: yn = xn

Each output value yn is exactly the same as the corresponding input value xn:

2. SIMPLE GAIN FILTER: yn = Kxn (K = constant) Amplifier or attenuator) This

simply applies a gain factor K to each input value:

3. PURE DELAY FILTER: yn = xn-1

The output value at time t = nh is simply the input at time t = (n-1)h, i.e. the
signal is delayed by time h:

4. TWO-TERM DIFFERENCE FILTER: yn = xn - xn-1

The output value at t = nh is equal to the difference between the current input xn
and the previous input xn-1:

5. TWO-TERM AVERAGE FILTER: yn = (xn + xn-1) / 2

The output is the average (arithmetic mean) of the current and previous input:
96

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

6. THREE-TERM AVERAGE FILTER: yn = (xn + xn-1 + xn-2) / 3

This is similar to the previous example, with the average being taken of the current
and two previous inputs.

7. CENTRAL DIFFERENCE FILTER: yn = (xn - xn-2) / 2

This is similar in its effect to example (4). The output is equal to half the change in
the input signal over the previous two sampling intervals:

ORDER OF A DIGITAL FILTER

The order of a digital filter can be defined as the number of previous inputs (stored in the processor's
memory) used to calculate the current output.
This is illustrated by the filters given as examples in the previous section.

Example (1): yn = xn
This is a zero order filter, since the current output yn depends only on the current
input xn and not on any previous inputs.

Example (2): yn = Kxn

The order of this filter is again zero, since no previous outputs are required to give
the current output value.

Example (3): yn = xn-1

This is a first order filter, as one previous input (xn-1) is required to calculate yn.
(Note that this filter is classed as first-order because it uses one previous input,
even though the current input is not used).

Example (4): yn = xn - xn-1

This is again a first order filter, since one previous input value is required to give
the current output.

Example (5): yn = (xn + xn-1) / 2

The order of this filter is again equal to 1 since it uses just one previous input
value.

Example (6): yn = (xn + xn-1 + xn-2) / 3

To compute the current output yn, two previous inputs (xn-1 and xn-2) are needed;
this is therefore a second-order filter.

Example (7): yn = (xn - xn-2) / 2

The filter order is again 2, since the processor must store two previous inputs in
order to compute the current output. This is unaffected by the absence of an explicit xn-1
term in the filter expression.

Q) For each of the following filters, state the order of the filter and identify the values of its
coefficients:
(a) yn = 2xn - xn-1 A) Order = 1: a0 = 2, a1 = -1
(b) yn = xn-2 B) Order = 2: a0 = 0, a1 = 0, a2 = 1
(c) yn = xn - 2xn-1 + 2xn-2 + xn-3 C) Order = 3: a0 = 1, a1 = -2, a2 = 2, a3 = 1
97

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

98

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

99

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

100

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

101

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

102

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

103

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

104

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

105

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

106

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

107

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

GLOSSARY:

FIR Filters:

In the Finite Impulse Response Filters the No.of. Impulses to be considered

for filtering are finite. There are no feed back Connections from the Output to the
Input. There are no Equivalent Structures of FIR filters in the Analog Regime.

Symmetric FIR Filters:

Symmetric FIR Filters have their Impulses that occur as the mirror image in
the first quadrant and second quadrant or Third quadrant and fourth quadrant or
both.

Anti Symmetric FIR Filters:

The Antisymmetric FIR Filters have their impulses that occur as the mirror
image in the first quadrant and third quadrant or second quadrant and Fourth

108

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

quadrant or both.
Linear Phase:
The FIR Filters are said to have linear in phase if the filter have the impulses
that increases according to the Time in digital domain.
Frequency Response:
The Frequency response of the Filter is the relationship between the angular
frequency and the Gain of the Filter.
Gibbs Phenomenon:
The abrupt truncation of Fourier series results in oscillation in both passband
and stop band. These oscillations are due to the slow convergence of the fourier
series. This is termed as Gibbs Phenomenon.
Windowing Technique:
To avoid the oscillations instead of truncating the fourier co-efficients we
are multiplying the fourier series with a finite weighing sequence called a window
which has non-zero values at the required interval and zero values for other
Elements.

109

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

UNIT – IV
FINITE WORDLENGTH EFFECTS
4.1 Number Representation
In digital signal processing, (B + 1)-bit fixed-point numbers are usually
represented as two’s- complement signed fractions in the format

bo b-ib-2 …… b-B
The number represented is then
X = -bo + b-i2- 1 + b-22 - 2 + ••• + b-B 2-B (3.1)
where bo is the sign bit and the number range is —1 <X < 1. The advantage of this
representation is that the product of two numbers in the range from — 1 to 1 is
another number in the same range. Floating-point numbers are represented as
X = (-1) s m2 c (3.2)
where s is the sign bit, m is the mantissa, and c is the characteristic or
exponent. To make the representation of a number unique, the mantissa is
normalized so that 0.5 <m < 1.
Although floating-point numbers are always represented in the form of (3.2), the
way in which this representation is actually stored in a machine may differ. Since
m > 0.5, it is not necessary to store the 2- 1 -weight bit of m, which is always set.
Therefore, in practice numbers are usually stored as
X = (-1)s(0.5 + f)2 c (3.3)
where f is an unsigned fraction, 0 <f < 0.5.
Most floating-point processors now use the IEEE Standard 754 32-bit floating-
point format for storing numbers. According to this standard the exponent is stored
as an unsigned integer p where
p = c + 126 (3.4)
Therefore, a number is stored as
X = (-1)s(0.5 + f )2p - 1 2 6 (3.5)
where s is the sign bit, f is a 23-b unsigned fraction in the range 0 <f < 0.5, and p
is an 8-b unsigned integer in the range 0 <p < 255. The total number of bits is 1 +
23 + 8 = 32. For example, in IEEE format 3/4 is written (-1)0 (0.5 + 0.25)2° so s =
0, p = 126, and f = 0.25. The value X = 0 is a unique case and is represented by
all bits zero (i.e., s = 0, f = 0, and p = 0). Although the 2-1-weight mantissa bit is
not actually stored, it does exist so the mantissa has 24 b plus a sign bit.

4.1.1 Fixed-Point Quantization Errors

In fixed-point arithmetic, a multiply doubles the number of significant bits. For
example, the product of the two 5-b numbers 0.0011 and 0.1001 is the 10-b number
00.000 110 11. The extra bit to the left of the decimal point can be discarded
without introducing any error. However, the least significant four of the remaining
bits must ultimately be discarded by some form of quantization so that the result
can be stored to 5 b for use in other calculations. In the example above this results
in 0.0010 (quantization by rounding) or 0.0001 (quantization by truncating). When
a sum of products calculation is performed, the quantization can be performed

110

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

either after each multiply or after all products have been summed with double-
length precision.
We will examine three types of fixed-point quantization—rounding, truncation,
and magnitude truncation. If X is an exact value, then the rounded value will be
denoted Q r (X), the truncated value Q t (X), and the magnitude truncated value
Q m t (X). If the quantized value has B bits to the right of the decimal point, the
quantization step size is
A = 2-B (3.6)
Since rounding selects the quantized value nearest the unquantized value, it gives a
value which is never more than ± A /2 away from the exact value. If we denote the
rounding error by
fr = Qr(X) - X (3.7)
then
AA
<f r < — (3.8)
2 - 2
Truncation simply discards the low-order bits, giving a quantized value that is
always less than or equal to the exact value so
- A < ft< 0 (3.9)
Magnitude truncation chooses the nearest quantized value that has a magnitude less
than or equal to the exact value so
—A
<fmt <A (3 10)
.
The error resulting from quantization can be modeled as a random variable
uniformly distributed over the appropriate error range. Therefore, calculations with
roundoff error can be considered error-free calculations that have been corrupted
by additive white noise. The mean of this noise for rounding is
1fA/2
m ( r — E{fr } = x/ frdfr — 0 (3.11)
A
J-A/2
where E{} represents the operation of taking the expected value of a random
variable. Similarly, the variance of the noise for rounding is

2 1 (A/2
2 2 A2
a€ — E{(fr - m€r) } — — (fr - m€r) dfr = — (3.12)
12
A
-A/2
Likewise, for truncation,
A
m E{f }
ft = t = -y
2 A2
2
2
a = E{(ft - mft)} = — (3.13)
m
fmt — E { f mt }— 0
and, for magnitude truncation

111

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

2 A2
2
a E{(f -m ) (3 14)
f-mt = mt m }=— .

4.1.2 Floating-Point Quantization Errors

With floating-point arithmetic it is necessary to quantize after both multiplications
and additions. The addition quantization arises because, prior to addition, the
mantissa of the smaller number in the sum is shifted right until the exponent of
both numbers is the same. In general, this gives a sum mantissa that is too long and
so must be quantized.
We will assume that quantization in floating-point arithmetic is performed by
rounding. Because of the exponent in floating-point arithmetic, it is the relative
error that is important. The relative error is defined as
Qr(X) - X er
e r = ----------- =— (3.15)
r
XX
Since X = (-1)sm2c, Qr(X) = (-1)s Qr(m)2c and
Q (m) - m e
r
er = ------------ =— (3.16)
mm
If the quantized mantissa has B bits to the right of the decimal point, | e| < A/2
where, as before, A = 2-B. Therefore, since 0.5 <m < 1,
|er I < A (3.17)
If we assume that e is uniformly distributed over the range from - A/2 to A/2 and
m is uniformly distributed over 0.5 to 1,
mSr = E \ — } = 0
m
a 2 = E {(—)2| = 2 f' A 4 dedm
er
I ' m/ J A J1/2J-a/2 m 2
A2
= — = (0.167)2-2B (3.18)
6
In practice, the distribution of m is not exactly uniform. Actual measurements of
roundoff noise in [1] suggested that
al r « 0.23A2 (3.19)
while a detailed theoretical and experimental analysis in [2] determined
a2 « 0.18A2 (3.20)
From (3.15) we can represent a quantized floating-point value in terms of the
unquantized value and the random variable er using
Qr(X) = X(1 + er) (3.21)
Therefore, the finite-precision product X1X2 and the sum X1 + X2 can be written
f IX1X2) = X1X2U + e r ) (3.22)
and

112

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

fl(X1 + X2) = (X1 + X2 )(1 + er) (3.23)

where e r is zero-mean with the variance of (3.20).

4.2 Roundoff Noise:

To determine the roundoff noise at the output of a digital filter we will assume that
the noise due to a quantization is stationary, white, and uncorrelated with the filter
input, output, and internal variables. This assumption is good if the filter input
changes from sample to sample in a sufficiently complex manner. It is not valid for
zero or constant inputs for which the effects of rounding are analyzed from a limit
cycle perspective.
To satisfy the assumption of a sufficiently complex input, roundoff noise in
digital filters is often calculated for the case of a zero-mean white noise filter input
signal x(n) of variance a 1. This simplifies calculation of the output roundoff noise
because expected values of the form E{x(n)x(n — k)} are zero for k = 0 and
give a2 when k = 0. This approach to analysis has been found to give estimates of
the output roundoff noise that are close to the noise actually observed for other
input signals.
Another assumption that will be made in calculating roundoff noise is that the
product of two quantization errors is zero. To justify this assumption, consider the
case of a 16-b fixed-point processor. In this case a quantization error is of the order
2—1 5 , while the product of two quantization errors is of the order 2 —3 0 , which is
negligible by comparison.
If a linear system with impulse response g(n) is excited by white noise with
mean m x and variance a2 , the output is noise of mean [3, pp.788-790]
TO
my = mx ^2g(n) (3.24)
n=—TO
and variance
TO
ay = al ^ g2 (n) (3.25)
n=—TO
Therefore, if g(n) is the impulse response from the point where a roundoff takes
place to the filter output, the contribution of that roundoff to the variance (mean-
square value) of the output roundoff noise is given by (3.25) with a 2 replaced with
the variance of the roundoff. If there is more than one source of roundoff error in
the filter, it is assumed that the errors are uncorrelated so the output noise variance
is simply the sum of the contributions from each source.

4.3 Roundoff Noise in FIR Filters:

The simplest case to analyze is a finite impulse response (FIR) filter realized via
—2B
1 2
al = (3.28)

113

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

the convolution summation

N—1
y(n) =Eh(k)x(n — k)
(3.26)
k=0
When fixed-point arithmetic is used and quantization is performed after each
multiply, the result of the N multiplies is N-times the quantization noise of a single
multiply. For example, rounding after each multiply gives, from (3.6) and (3.12),
an output noise variance of
2 2 —2B
2
a = N— (3.27)
Virtually all digital signal processor integrated circuits contain one or more double-
length accumulator registers which permit the sum-of-products in (3.26) to be
accumulated without quantization. In this case only a single quantization is
necessary following the summation and
For the floating-point roundoff noise case we will consider (3.26) for N = 4 and
then generalize the result to other values of N. The finite-precision output can be
written as the exact output plus an error term e(n). Thus,
y(n) + e(n) = ({[h(0)x(n)[1 + E1(n)]
+ h(1)x(n -1)[1 + £2(n)]][1 + S3(n)]
+ h(2)x(n -2)[1 + £4(n)]}{1 + s5(n)}
+ h(3)x(n - 3)[1 + £6(n)])[1 + £j(n)] (3.29)
In (3.29), £1(n) represents the errorin the first product, £2(n) the
error in the second product, £3(n)
the error in the firstaddition, etc.Notice that it has been assumed that the products
are summed in
the order implied by the summation of (3.26).
Expanding (3.29), ignoring products of error terms, and recognizing y(n) gives
e(n) = h(0)x(n)[£1 (n) + £3(n) + £$(n) + £i(n)]
+ h(1)x(n-1)[£2(n) + £3(n) + £5(n) + £j(n)]
+ h(2)x(n-2)[£4(n) + £5(n) + £i(n)]
+ h(3)x(n- 3)[£6(n)+ £j(n)]
(3.30)
Assuming that the input is white noise of variance a^ so that E{x(n)x(n - k)} is
zero for k = 0, and assuming that the errors are uncorrelated,
E{e2(n)} = [4h2(0)+ 4h2(1) + 3h2(2) + 2h2(3)]a2a2 (3.31)
In general, for any N,
N-1

aO = E{e 2 (n)} = Nh2(0) +J2 (N + 1 - a°a° r (3.32)

k)h2(k)

k=1

114

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Notice that if the order of summation of the product terms in the convolution
summation is changed, then the order in which the h(k)’s appear in (3.32)
changes. If the order is changed so that the h(k)with smallest magnitude is first,
followed by the next smallest, etc., then the roundoff noise variance is
minimized. However, performing the convolution summation in nonsequential
order greatly complicates data indexing and so may not be worth the reduction
obtained in roundoff noise.

4.4 Roundoff Noise in Fixed-Point IIR Filters

To determine the roundoff noise of a fixed-point infinite impulse response (IIR)

filter realization, consider a causal first-order filter with impulse response
h(n) = anu(n) (3.33)
realized by the difference equation
y(n) = ay(n - 1) + x(n) (3.34)
Due to roundoff error, the output actually obtained is
y(n) = Q{ay(n - 1) + x(n)} = ay(n - 1) + x(n) + e(n) (3.35)

115

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

where e(n) is a random roundoff noise sequence. Since e(n) is injected at the
same point as the input, it propagates througha system with impulse
response h(n). Therefore, forfixed-point arithmetic
with rounding,the outputroundoff noise variance from (3.6), (3.12), (3.25),
and (3.33) is
2 2
2 A “ 2 A “ 2n 2-2B 1
2 2 2n
a = — > h (n) = — > a = ---------------------------------- (3.36)
o ()
12 ^ 12 ^ 12 1 - a 2
n=—<x n=0
With fixed-point arithmetic there is the possibility of overflow following
addition. To avoid overflow it is necessary to restrict the input signal amplitude.
This can be accomplished by either placing a scaling multiplier at the filter
input or by simply limiting the maximum input signal amplitude. Consider the
case of the first-order filter of (3.34). The transfer function of this filter is
,v, Y(e j m ) 1
jm
H(e ) = . = (3.37)
X(eJm) eJm — a
so
\H(e j m )\ 2 = --- - —1 ------------------------------- (3.38)
1 + a 2 — 2a cos,)
and
,7, 1
|H(e^)|max = ---- — (3.39)
1 — \a\
The peak gain of the filter is 1/(1 — \a\) so limiting input signal amplitudes to
\x(n)\ < 1 — \ a | will make overflows unlikely.
An expression for the output roundoff noise-to-signal ratio can easily be
obtained for the case where the filter input is white noise, uniformly distributed
over the interval from —(1 — \ a \) to (1 — \ a \) [4,5]. In this case
? 1 f 1 —\ a \ ? 1 ?
a x dx (1 — a (3 40)
x = 21—a) =3 \ \ ) .
2(1 — a ) 3
\ \ J — (1 — \a \)
so, from (3.25),
2 1 (1 — \a \)2
2
ay = 3 , 2 (3'41)
y 2
31— a
Combining (3.36) and (3.41) then gives
0^ = ( 2 _—l^\ (3^0L^ = 2—B (3
42)
a2 \12 1 — a 2 )\ (1 — \a \) 2 ) 12 (1— \ a \ ) 2 (
. )
Notice that the noise-to-signal ratio increases without bound as \ a \ ^ 1.
Similar results can be obtained for the case of the causal second-order filter

116

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

realized by the difference equation

y(n) = 2r cos(0)y(n — 1) — r2y(n — 2) + x(n)
(3.43)
This filter has complex-conjugate poles at re± j 0 and impulse response
h(n) =r n sin[(n + 1)0] u(n)
--------------------------- (3.44)
sin (0)
Due to roundoff error, the output actually obtained is
y(n) = 2r cos(0)y(n — 1) — r2y(n — 2) + x(n) + e(n) (3.45)

117

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

There are two noise sources contributing to e(n) if quantization is performed after
each multiply, and there is one noise source if quantization is performed after
summation. Since
2 1
1 — r 2
(1 + r 2 2
) — 4r2 (3.46
n= )
— oo cos2 (9)
the output roundoff
noise is
2—2B1 + 1
a2 = V 2 (3.47
r 12 1— r 2 (1 + r 2 )2 — 4r2
2 )
where V = 1 for quantization cos
after(9)
summation, and V = 2 for quantization
after each multiply. To obtain an output noise-to-signal ratio we note that

H(e j w ) 1
(3.48
= 1 — 2r cos(9)e — j m +
)
r2e—j2m
and, using the approach of
[6],
iH(emmax = 1
(3.49
2 sat ( c o s ( 9 ) ^ — cos(9) 2 2
sin 2
4r + )
(9)
wher
e I >1
sat (i) — 1<I<1 (3.50
= )
Following the same approach as for the first-order
case then gives
2 2 — 2 B 1+r 2 3
V 12 1 — r 2 (1 + r 2 )2 — 4r2
y
cos2 (9) 1
X 2 (^cos 2 2 (3.51
4r sat (9) cos( + 1— 2 sin(9) )
) 9) r
Figure3.1 is a contour plot showing the noise-to-signal ratio 2rof (3.51) for v = 1
in units of the noise variance of a single quantization, 2—2 B /12. The plot is
symmetrical about 9 = 90°, so only the range from 0° to 90° is shown. Notice
that as r ^ 1, the roundoff noise increases without bound. Also notice that the
noise increases as 9 ^ 0°.
It is possible to design state-space filter realizations that minimize fixed-point
roundoff noise [7] - [10]. Depending on the transfer function being realized, these
structures may provide a roundoff noise level that is orders-of-magnitude lower
than for a nonoptimal realization. The price paid for this reduction in roundoff
noise is an increase in the number of computations required to implement the
filter. For an Nth-order filter the increase is from roughly 2N multiplies for a
direct form realization to roughly (N + 1)2 for an optimal realization. However,
if the filter is realized by the parallel or cascade connection of first- and second-
order optimal subfilters, the increase is only to about 4N multiplies. Furthermore,
near-optimal realizations exist that increase the number of multiplies to only
about 3N [10].
118

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

0.010.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99
Normalized fixed-point roundoff noise variance.

4.5 Limit Cycle Oscillations:

A limit cycle, sometimes referred to as a multiplier roundoff limit cycle, is a low-
level oscillation that can exist in an otherwise stable filter as a result of the
nonlinearity associated with rounding (or truncating) internal filter calculations
[11]. Limit cycles require recursion to exist and do not occur in nonrecursive FIR
filters. As an example of a limit cycle, consider the second-order filter realized by
7 5
2 x
y(n) = Qr{ ^y(n —1) — 8y(n — ) + (n)
where Q r {} represents quantization by rounding. This is stable filter with poles
at 0.4375 ± j0.6585. Consider the implementation of this filter with 4-b (3-b and
a sign bit) two’s complement fixed-point arithmetic, zero initial conditions (y(—
1) = y(—2) = 0), and an input sequence x(n) = |S(n), where S(n) is the unit
impulse or unit sample. The following sequence is obtained;

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Notice that while the input is zero except for the first sample, the output oscillates
with amplitude 1/8 and period 6.
Limit cycles are primarily of concern in fixed-point recursive filters. As long as
floating-point filters are realized as the parallel or cascade connection of first- and
second-order subfilters, limit cycles will generally not be a problem since limit
cycles are practically not observable in first- and second-order systems
implemented with 32-b floating-point arithmetic [12]. It has been shown that such
systems must have an extremely small margin of stability for limit cycles to exist
at anything other than underflow levels, which are at an amplitude of less than
10 — 3 8 [12]. There are at least three ways of dealing with limit cycles when fixed-
point arithmetic is used. One is to determine a bound on the maximum limit cycle
amplitude, expressed as an integral number of quantization steps [13]. It is then
possible to choose a word length that makes the limit cycle amplitude acceptably
low. Alternately, limit cycles can be prevented by randomly rounding calculations
up or down [14]. However, this approach is complicated to implement. The third
approach is to properly choose the filter realization structure and then quantize
the filter calculations using magnitude truncation [15,16]. This approach has the
disadvantage of producing more roundoff noise than truncation or rounding [see
(3.12)—(3.14)].

4.6 Overflow Oscillations:

With fixed-point arithmetic it is possible for filter calculations to overflow. This
happens when two numbers of the same sign add to give a value having
magnitude greater than one. Since numbers with magnitude greater than one are
not representable, the result overflows. For example, the two’s complement
numbers 0.101 (5/8) and 0.100 (4/8) add to give 1.001 which is the two’s
complement representation of -7/8.
The overflow characteristic of two’s complement arithmetic can be represented
as R{} where
X - X>1
R{X}
For the example just considered, R{9/8}
2 X = —7/8.
-1 < X < 1 (3.71
An overflow oscillation, = sometimes
X+2 also Xreferred
<-1 to as an adder overflow )
limit cycle, is a high- level oscillation that can exist in an otherwise stable
fixed-point filter due to the gross nonlinearity associated with the overflow of
internal filter calculations [17]. Like limit cycles, overflow oscillations require
recursion to exist and do not occur in nonrecursive FIR filters. Overflow
oscillations also do not occur with floating-point arithmetic due to the virtual
impossibility of overflow.
As an example of an overflow oscillation, once again consider the filter of
(3.69) with 4-b fixed-point two’s complement arithmetic and with the two’s
complement overflow characteristic of (3.71):

75
y(n) = Qr\R (3.72)
8y(n -1) - 8y(n - 2) + x(n)

[Type text]
Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net
In this case we apply the
input

35
x(n) = -4 (n) - ^&(n -1)
&

3 5 ,,
(3.73)
0, 0, ■
4 8
s to scale the filter calculations so as to render overflow impossible. However,
this may unacceptably restrict the filter dynamic range. Another method is to
force completed sums-of- products to saturate at ±1, rather than overflowing
[18,19]. It is important to saturate only the completed sum, since intermediate
overflows in two’s complement arithmetic do not affect the accuracy of the final
result. Most fixed-point digital signal processors provide for automatic saturation
of completed sums if their saturation arithmetic feature is enabled. Yet
another way to avoid overflow oscillations is to use a filter structure for which
any internal filter transient is guaranteed to decay to zero [20]. Such structures are
desirable anyway, since they tend to have low roundoff noise and be insensitive
to coefficient quantization [21].

4.7 Coefficient Quantization Error:

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

0
0.25 0.50 0.75 1.00
Re Z
FIGURE: Realizable pole locations for the
difference equation of (3.76).

The sparseness of realizable pole locations near z = ± 1 will result in a large coefficient
quantization error for poles in this region.
Figure3.4 gives an alternative structure to (3.77) for realizing the transfer function of (3.76).
Notice that quantizing the coefficients of this structure corresponds to quantizing X r and Xi.
As shown in Fig.3.5 from [5], this results in a uniform grid of realizable pole locations.
Therefore, large coefficient quantization errors are avoided for all pole locations.
It is well established that filter structures with low roundoff noise tend to be robust to
coefficient quantization, and visa versa [22]- [24]. For this reason, the uniform grid
structure of Fig.3.4 is also popular because of its low roundoff noise. Likewise, the low-
noise realizations of [7]- [10] can be expected to be relatively insensitive to coefficient
quantization, and digital wave filters and lattice filters that are derived from low-sensitivity
analog structures tend to have not only low coefficient sensitivity, but also low roundoff
noise [25,26].
It is well known that in a high-order polynomial with clustered roots, the root location is a
very sensitive function of the polynomial coefficients. Therefore, filter poles and zeros can
be much more accurately controlled if higher order filters are realized by breaking them up
into the parallel or cascade connection of first- and second-order subfilters. One exception
to this rule is the case of linear-phase FIR filters in which the symmetry of the polynomial
coefficients and the spacing of the filter zeros around the unit circle usually permits an
acceptable direct realization using the convolution summation.
Given a filter structure it is necessary to assign the ideal pole and zero locations to the
realizable locations. This is generally done by simplyrounding or truncatingthe filter
coefficients to the available number of bits, or by assigning the ideal pole and zero locations
to the nearest realizable locations. A more complicated alternative is to consider the original
filter design problem as a problem in discrete

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

FIGURE 3.4: Alternate realization

structure.

FIGURE 3.5: Realizable pole locations for the

alternate realization structure.

optimization, and choose the realizable pole and zero locations that give the best
approximation to the desired filter response [27]- [30].

4.8 Realization Considerations:

Linear-phase FIR digital filters can generally be implemented with acceptable coefficient
quantization sensitivity using the direct convolution sum method. When implemented in
this way on a digital signal processor, fixed-point arithmetic is not only acceptable but may
actually be preferable to floating-point arithmetic. Virtually all fixed-point digital signal
processors accumulate a sum of products in a double-length accumulator. This means that
only a single quantization is necessary to compute an output. Floating-point arithmetic, on

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

the other hand, requires a quantization after every multiply and after every add in the
convolution summation. With 32-b floating-point arithmetic these quantizations introduce a
small enough error to be insignificant for many applications.
When realizing IIR filters, either a parallel or cascade connection of first- and second-order
subfilters is almost always preferable to a high-order direct-form realization. With the
availability of very low-cost floating-point digital signal processors, like the Texas
Instruments TMS320C32, it is highly recommended that floating-point arithmetic be used
for IIR filters. Floating-point arithmetic simultaneously eliminates most concerns regarding
scaling, limit cycles, and overflow oscillations. Regardless of the arithmetic employed, a
low roundoff noise structure should be used for the second- order sections. Good choices
are given in [2] and [10]. Recall that realizations with low fixed-point roundoff noise also
have low floating-point roundoff noise. The use of a low roundoff noise structure for the
second-order sections also tends to give a realization with low coefficient quantization
sensitivity. First-order sections are not as critical in determining the roundoff noise and
coefficient sensitivity of a realization, and so can generally be implemented with a simple
direct form structure.
GLOSSARY:
Quantization:
Total number of bits in x is reduced by using two methods namely Truncation and Rounding. These are
known as quantization Processes.
Input Quantization Error:
The Quantized signal are stored in a b bit register but for nearest values the same digital equivalent may be
represented. This is termed as Input Quantization Error.
Product Quantization Error:
The Multiplication of a b bit number with another b bit number results in a 2b bit number but it should be
stored in a b bit register. This is termed as Product Quantization Error.
Co-efficient Quantization Error:
The Analog to Digital mapping of signals due to the Analog Co-efficient Quantization results in error due
to the Fact that the stable poles marked at the edge of the jΩ axis may be marked as an unstable pole in the
digital domain.
Limit Cycle Oscillations:
If the input is made zero, the output should be made zero but there is an error occur due to the quantization
effect that the system oscillates at a certain band of values.
Overflow limit Cycle oscillations:
Overflow error occurs in addition due to the fact that the sum of two numbers may result in overflow. To
avoid overflow error saturation arithmetic is used.
Dead band:
The range of frequencies between which the system oscillates is termed as Deadband of the Filter. It may
have a fixed positive value or it may oscillate between a positive and negative value.
Signal scaling:
The inputs of the summer is to be scaled first before execution of the addition operation to find for any
possibility of overflow to be occurred after addition. The scaling factor s0 is multiplied with the inputs to
avoid overflow.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

UNIT V
APPLICATIONS OF DSP
PRE REQUISITE DISCUSSION:
The time domain waveform is transformed to the frequency domain using a filter bank. The strength of
each frequency band is analyzed and quantized based on how much effect they have on the perceived
decompressed signal.

5.1. SPEECH RECOGNITION:

Basic block diagram of a speech recognition system is shown in Fig 1

1. In speech recognition system using microphone one can input speech or voice. The analog
speech signal is converted to digital speech signal by speech digitizer. Such digital signal is
called digitized speech.

2. The digitized speech is processed by DSP system. The significant features of speech such
as its formats, energy, linear prediction coefficients are extracted. The template of this
extracted features are compared with the standard
reference templates. The closed matched template is considered as the recognized
word.

3. Voice operated consumer products like TV, VCR, Radio, lights, fans and voice operated
telephone dialing are examples of DSP based speech recognized devices.

Impulse Voiced
Train Synthetic
Generator speech
Time
× varying
digital filter
Random
number
generator Unvoiced

5.2. LINEAR PREDICTION OF SPEECH SYNTHESIS

Fig shows block diagram of speech synthesizer using linear prediction.

1. For voiced sound, pulse generator is selected as signal source while for unvoiced sounds
noise generator is selected as signal source.

2. The linear prediction coefficients are used as coefficients of digital filter. Depending upon these
coefficients , the signal is passed and filtered by the digital filter.

3. The low pass filter removes high frequency noise if any from the synthesized speech. Because
of linear phase characteristic FIR filters are mostly used as digital filters.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Pitch Period

Pulse Voiced
Generator
Synthetic
speech
Digital Time
filter varying
digital
White filter
Noise
generator Unvoiced

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Filter Coefficients

5.3. SOUND PROCESSING:

1. In sound processing application, Music compression(MP3) is achieved by
converting the time domain signal to the frequency domain then removing frequencies which
are no audible.

2. The time domain waveform is transformed to the frequency domain using a filter bank. The
strength of each frequency band is analyzed and quantized based on how much effect they
have on the perceived decompressed signal.

3. The DSP processor is also used in digital video disk (DVD) which uses MPEG-2
compression, Web video content application like Intel Indeo, real audio.

4. Sound synthesis and manipulation, filtering, distortion, stretching effects are also done by DSP
processor. ADC and DAC are used in signal generation and recording.

5. 4. ECHO CANCELLATION

In the telephone network, the subscribers are connected to telephone exchange by two wire circuit. The
exchanges are connected by four wire circuit. The two wire circuit is bidirectional and carries signal in
both the directions. The four wire circuit has separate paths for transmission and reception. The hybrid
coil at the exchange provides the interface between two wire and four wire circuit which also provides
impedance matching between two wire and four wire circuits. Hence there are no echo or reflections on
the lines. But this impedance matching is not perfect because it is length dependent. Hence
for echo cancellation, DSP techniques are used as follows.
1. An DSP based acoustic echo canceller works in the following fashion: it records the sound
going to the loudspeaker and substract it from the signal coming from
the microphone. The sound going through the echo-loop is transformed and delayed,
and noise is added, which complicate the substraction process.

2. Let be the input signal going to the loudspeaker; let be the signal picked up by the
microphone, which will be called the desired signal. The signal after

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

substraction will be called the error signal and will be denoted by . The adaptive filter will
try to identify the equivalent filter seen by the system from the loudspeaker to the
microphone, which is the transfer function of the room the loudpeaker and microphone are in.

3. This transfer function will depend heavily on the physical characteristics of the environment. In
broad terms, a small room with absorbing walls will origninate just a few, first order reflections
so that its transfer function will have a short impulse response. On the other hand, large rooms
with reflecting walls will have a transfer function whose impulse response decays slowly in
time, so that echo cancellation will be much more difficult.

5. 5 VIBRATION ANALYSIS:

1. Normally machines such as motor, ball bearing etc systems vibrate depending upon the speed of
their movements.

2. In order to detect fault in the system spectrum analysis can be performed. It shows fixed
frequency pattern depending upon the vibrations. If there is fault in the machine, the
predetermined spectrum is changes. There are new frequencies introduced in the spectrum
representing fault.

3. This spectrum analysis can be performed by DSP system. The DSP system can also be used to
monitor other parameters of the machine simultaneously.

Entire Z Plane except

Z=0

Entire Z Plane except

Z=∞

Entire z Plane except Z

=0 & Z=∞

Visit : www.EasyEngineeering.net
Visit : www.EasyEngineeering.net

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

5.6. Multistage Implementation of Digital Filters:

In some applications we want to design filters where the bandwidth is just a small fraction of
the overall sampling rate. For example, suppose we want to design a lowpass filter with
bandwidth of the order of a few hertz and a sampling frequency of the order of several kilohertz.
This filter would require a very sharp transition region in the digital frequency a>, thus
requiring a high-complexity filter.
Example <
As an example of application, suppose you want to design a Filter with
the fallowing specifications:
Passband Fp = 450 Hz
Stopband F s =500 Hz
Sampling frequency F s ~96 kHz
Notice that the stopband is several orders of magnitude smaller than the
sampling frequency. This leads to a filter with a very short transition
region of high complexity. In
Speech signals
© From prehistory to the new media of the future, speech has been and will be a
primary form of communication between humans.
© Nevertheless, there often occur conditions under which we measure and then
transform the speech to another form, speech signal, in order to enhance our ability
to communicate.
© The speech signal is extended, through technological media such as telephony,
movies, radio, television, and now Internet. This trend reflects the primacy of speech
communication in human psychology.
© “Speech will become the next major trend in the personal computer market in the near
future.”

5.7.Speech signal processing:

© The topic of speech signal processing can be loosely defined as the manipulation of
sampled speech signals by a digital processor to obtain a new signal with some desired
properties.
Speech signal processing is a diverse field that relies on knowledge of language at the
levels of Signal processing
Acoustics (P¥)
Phonetics ( ^ ^ ^ ) Language-independent
Phonology ( ^ ^ )
Morphology ( i ^ ^ ^ )
Syntax ( ^ , £ ) Language-dependent
Semantics (\%X¥)
Pragmatics ( i f , f f l ^ )
7 layers for describing speech From Speech to Speech Signal, in terms of
Digital Signal Processing
At-*
■ Acoustic (and perceptual) features
{traits)
- fundamental freouency (FO)
(pitch)
- amplitude (loudness)
- spectrum (timber)
[Type text]
Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

It is based on the fact that

- Most of energy between 20 Hz to about 7KHz ,
- Human ear sensitive to energy between 50 Hz and 4KHz
© In terms of acoustic or perceptual, above features are considered.
© From Speech to Speech Signal, in terms of Phonetics (Speech production), the
digital model of Speech Signal will be discussed in Chapter 2.
© Motivation of converting speech to digital signals:

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

pitch,
Speech coding, A PC and SBC fine structure
Adaptive predictive coding (APC) is a technique used for speech coding, that is data
compression of spccch signals
APC assumes that the input
Typic speech signal is repetitive with a
al period significantly longer than
Voice the average frequency content.
d Two predictors arc used in APC.
speec The high frequency components
h (up to 4 kHz) are estimated using
a 'spectral’ or 'formant’ prcdictor and the low frequency components (50-200 Hz) by a ‘pitch’
or ‘fine structure’ prcdictor (see figure 7.4). The spcctral estimator may he of order 1- 4 and the
pitch estimator about order 10. The low-frequency components of the spccch signal are due to
the movement of the tongue, chin and spectral
envelope, formants

Figure 7.4 Encoder for adaptive, predictive coding of speech

signals. The decoder is mainly a mirrored version of the encoder

The high-frequency components originate from the vocal chords and the noise-like sounds (like
in ‘s’) produced in the front of the mouth.
The output signal y(n)together with the predictor parameters, obtained adaptively in the
encoder, are transmitted to the decoder, where the spcech signal is reconstructed. The decoder
has the same structure as the encoder but the predictors arc not adaptive and arc invoked in the
reverse order. The prediction parameters are adapted for blocks of data corresponding to for
instance 20 ms time periods.
A PC' is used for coding spcech at 9.6 and 16 kbits/s. The algorithm works well in noisy
environments, but unfortunately the quality of the processed speech is not as good as for other
methods like CELP described below.

5..8 Subband Coding:

Another coding method is sub-band coding (SBC) (see Figure 7.5) which belongs to the
class of waveform coding methods, in which the frequency domain properties of the input
signal arc utilized to achieve data compression.
The basic idea is that the input speech signal is split into sub-bands using band-pass filters.
The sub-band signals arc then encoded using ADPCM or other techniques. In this way, the
available data transmission capacity can be allocated between bands according to pcrccptual
criteria, enhancing the speech quality as pcrceivcd by listeners. A sub-band that is more
‘important’ from the human listening point of view can be allocated more bits in the data
stream, while less important sub-bands will use fewer bits.
A typical setup for a sub-band codcr would be a bank of N(digital) bandpass filters followed

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

by dccimators, encoders (for instance ADPCM) and a multiplexer combining the data bits
coming from the sub-band channels. The output of the multiplexer is then transmitted to the
sub-band dccodcr having a demultiplexer splitting the multiplexed data stream back into Nsub-
band channels. Every sub-band channel has a dccodcr (for instance ADPCM), followed by an
interpolator and a band-pass filter. Finally, the outputs of the band-pass filters are summed and
a reconstructed output signal results.
Sub-band coding is commonly used at bit rates between 9.6 kbits/s and 32 kbits/s and
performs quite well. The complexity of the system may however be considerable if the number
of sub-bands is large. The design of the band-pass filters is also a critical topic when working
with sub-band coding systems.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Figure 7.5 An example of a sub-band coding system

Vocoders and LPC

In the methods described above (APC, SBC and ADPCM),
speech signal applications have been used as examples. By
modifying the structure and parameters of the predictors and
filters, the algorithms may also be used for other signal types.
The main objective was to achieve a reproduction that was as
faithful as possible to the original signal. Data compression was
possible by removing redundancy in the time or frequency
domain.
The class of vocoders (also called source coders) is a special class of
data compression devices aimed only at spcech signals. The input signal
is analysed and described in terms of speech model parameters. These
parameters are then used to synthesize a voice pattern having an
acceptable level of perceptual quality. Hence, waveform accuracy is not
the main goal, as in the previous methods discussed.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

periodic
pitch excitation

9+V
synthetic
noise speech

voiced/unvoiced
Figure 7.6 The LPC model

The first vocoder was designed by H. Dudley in the 1930s and demonstrated at
the ‘New York Fair’ in 1939* Vocoders have become popular as they achieve
reasonably good speech quality at low data rates, from 2A kbits/s to 9,6 kbits/s.
There arc many types of vocoders (Marvcn and Ewers, 1993), some of the most
common techniques will be briefly presented below.
Most vocoders rely on a few basic principles. Firstly, the characteristics of the
spccch signal is assumed to be fairly constant over a time of approximately 20
ms, hcncc most signal processing is performed on (overlapping) data blocks of 20
40 ms length. Secondly, the spccch model consists of a time varying filter
corresponding to the acoustic properties of the mouth and an excitation signal.
The cxeilalion signal is cither a periodic waveform, as crcatcd by the vocal
chords, or a random noise signal for production of ‘unvoiced' sounds, for
example ‘s’ and T. The filter parameters and excitation parameters arc assumed
to be independent of each other and are commonly coded separately.
Linear predictive coding (LPC) is a popular method, which has however
been replaced by newer approaches in many applications. LPC works exceed-
ingly well at low bit rates and the LPC parameters contain sufficient information
of the spccch signal to be used in spccch recognition applications. The LPC
model is shown in Figure 7*6.
LPC is basically an autn-regressive model (sec Chapter 5) and the vocal tract
is modelled as a time-varying all-pole filter (HR filter) having the transfer
function H(z)
(7*17)
-k
k=I
where p is the order of the filter. The excitation signal *?(«), being either noise or
a periodic waveform, is fed to the filter via a variable gain factor G. The output
signal can be expressed in the time domain as

y(n) ~ Ge(ri) - a , y ( n - 1) - a 2 y ( n - 2 ) %y{n-p) ( 1 . IK)

The output sample at time n is a linear combination of p previous samples

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

and the excitation signal (linear predictive coding). The filter coefficients a k arc
time varying.
The model above describes how to synthesize the speech given the pitch
information (if noise or pet iodic excitation should be used), the gain and the
filter parameters. These parameters must be determined by the cncoder or the
analyser, taking the original spccch signal x(n) as input.
The analyser windows the spccch signal in blocks of 20-40 ms. usually with a
Hamming window (see Chapter 5). These blocks or ‘frames’ arc repeated every
10—30 ms, hence there is a certain overlap in time. Every frame is then analysed
with respect to the parameters mentioned above.
Firstly, the pitch frequency is determined. This also tells whether we arc
dealing with a voiced or unvoiccd spccch signal. This is a crucial part of the
system and many pitch detection algorithms have been proposed. If the segment
of the spccch signal is voiced and has a dear periodicity or if it is unvoiced and
not pet iodic, things arc quite easy* Segments having properties in between these
two extremes are difficult to analyse. No algorithm has been found so far that is
1
perfect* for all listeners.
Now, the second step of the analyser is to determine the gain and the filter
parameters. This is done by estimating the spccch signal using an adaptive
predictor. The predictor has the same structure and order as the filter in the
synthesizer, Hencc, the output of the predictor is
- i ( n ) — - tf] jt(/7— 1) — a 2 x ( n — 2) — . . . — OpX(n—p) (7-19)
where i(rt) is the predicted input spcech signal and jc(rt) is the actual input signal.
The filter coefficients a k are determined by minimizing the square error

This can be done in different ways, cither by calculating the auto-corrc- lation
coefficients and solving the Yule Walker equations (see Chapter 5) or by using
some recursive, adaptive filter approach (see Chapter 3),
So, for every frame, all the parameters above arc determined and irans- mittcd to
the synthesiser, where a synthetic copy of the spccch is generated.
An improved version of LPC is residual excited linear prediction (RELP).
Let us take a closer look at the error or residual signal r(fi) resulting from the
prediction in the analyser (equation (7.19)). The residual signal (wc arc try ing to
minimize) can be expressed as
r(n)= *(«) -i(rt)
= jf(rt) + a^x(n— 1) + a 2 x(n— 2) -h *.. + a p x(ft—p) <7-21)
From this it is straightforward to find out that the corresponding expression using
the z-transforms is

Hcncc, the prcdictor can be regarded as an ‘inverse’ filter to the LPC model filter.
If we now pass this residual signal to the synthesizer and use it to excite the LPC
filter, that is E(z) - R(z), instead of using the noise or periodic waveform
sources we get
Y ( z ) = E ( z ) H ( z ) = R ( z ) H ( z ) = X ( z ) H ~ \ z ) H ( z ) = X ( z ) (7.23)
In the ideal case, we would hence get the original speech signal back. When
minimizing the variance of the residual signal (equation (7.20)), we gathered as
much information about the spccch signal as possible using this model in the
filter coefficients a k . The residual signal contains the remaining information. If
Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

the model is well suited for the signal type (speech signal), the residual signal is
close to white noise, having a flat spectrum. In such a case we can get away with
coding only a small range of frequencies, for instance 0-1 kHz of the residual
signal. At the synthesizer, this baseband is then repeated to generate higher
frequencies. This signal is used to excite the LPC filter
Vocoders using RELP are used with transmission rates of 9.6 kbits/s. The
advantage of RELP is a better speech quality compared to LPC for the same bit
rate. However, the implementation is more computationally demanding.
Another possible extension of the original LPC approach is to use multipulse
excited linear predictive coding (MLPC). This extension is an attempt to make
the synthesized spcech less ‘mechanical’, by using a number of different pitches
of the excitation pulses rather than only the two (periodic and noise) used by
standard LPC.
The MLPC algorithm sequentially detects k pitches in a speech signal. As soon
as one pitch is found it is subtracted from the signal and detection starts over
again, looking for the next pitch. Pitch information detection is a hard task and
the complexity of the required algorithms is often considerable. MLPC however
offers a better spcech quality than LPC for a given bit rate and is used in systems
working with 4.S-9.6 kbits/s.
Yet another extension of LPC is the code excited linear prediction (CELP).
The main feature of the CELP compared to LPC is the way in which the filter
coefficients are handled. Assume that we have a standard LPC system, with a
filter of the order p. If every coefficient a k requires N bits, we need to transmit
N-p bits per frame for the filter parameters only. This approach is all right if all
combinations of filter coefficients arc equally probable. This is however not the
case. Some combinations of coefficients are very probable, while others may
never occur. In CELP, the coefficient combinations arc represented by p
dimensional vectors. Using vector quantization techniques, the most probable
vectors are determined. Each of these vectors are assigned an index and stored in
a codebook. Both the analyser and synthesizer of course have identical copies of
the codebook, typically containing 256-512 vectors. Hcncc, instead of
transmitting N-p bits per frame for the filter parameters only 8-9 bits arc needed.
This method offers high-quality spcech at low-bit rates but requires consid-
erable computing power to be able to store and match the incoming spcech to the
‘standard’ sounds stored in the codebook. This is of course especially true if the
codebook is large. Speech quality degrades as the codebook size decreases.
Most CELP systems do not perform well with respect to higher frequency
components of the spccch signal at low hit rates. This is countcractcd in
There is also a variant of CELP called vector sum excited linear prediction
(VSELP). The main difference between CELP and VSELP is the way the
codebook is organized. Further, since VSELP uses fixed point arithmetic
algorithms, it is possible to implement using cheaper DSP chips than
Adaptive Filters
The signal degradation in some physical systems is time varying, unknown, or possibly
both. For example,consider a high-speed modem for transmitting and receiving data over
telephone channels. It employs a filter called a channel equalizer to compensate for the channel
distortion. Since the dial-up communication channels have different and time-varying
characteristics on each connection, the equalizer must be an adaptive filter.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

5.9 Adaptive Filter:

Adaptive filters modify their characteristics to achieve certain objectives by
automatically updating their coefficients. Many adaptive filter structures and adaptation
algorithms have been developed for different applications. This chapter presents the most widely
used adaptive filters based on the FIR filter with the least-mean-square (LMS) algorithm. These
adaptive filters are relatively simple to design and implement. They are well understood with
regard to stability, convergence speed, steady-state performance, and finite-precision effects.
Introduction to Adaptive Filtering
An adaptive filter consists of two distinct parts - a digital filter to perform the desired filtering,
and an adaptive algorithm to adjust the coefficients (or weights) of the filter. A general form of
adaptive filter is illustrated in Figure 7.1, where d(n) is a desired (or primary input) signal, y(n)
is the output of a digital filter driven by a reference input signal x(n), and an error signal e(n) is
the difference between d(n) and y(n). The adaptive algorithm adjusts the filter coefficients to
minimize the mean-square value of e(n). Therefore, the filter weights are updated so that the
error is progressively minimized on a sample-bysample basis.
In general, there are two types of digital filters that can be used for adaptive filtering: FIR and
IIR filters. The FIR filter is always stable and can provide a linear-phase response. On the other
hand, the IIR

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

filter involves both zeros and poles. Unless they are properly controlled, the poles in the filter
may move outside the unit circle and result in an unstable system during the adaptation of
coefficients. Thus, the adaptive FIR filter is widely used for practical real-time applications.
This chapter focuses on the class of adaptive FIR filters.
The most widely used adaptive FIR filter is depicted in Figure 7.2. The filter output signal
is computed

(7.13)

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

, where the filter coefficients wl (n) are time varying and updated by the adaptive algorithms that will be
discussed next.
We define the input vector at time n as
x(n) = [x(n)x(n - 1) . . . x(n - L + 1)]T , (7.14)
and the weight vector at time n as
w(n) = [w0(n)w1(n) . . . wL-1(n)]T . (7.15)
Equation (7.13) can be expressed in vector form as
y(n) = wT (n)x(n) = xT (n)w(n). (7.16)
The filter outputy(n) is compared with the desired d(n) to obtain the error signal e(n) =
d(n) - y(n) = d(n) - wT (n)x(n). (7.17)
Our objective is to determine the weight vector w(n) to minimize the predetermined performance (or cost)
function.
Performance Function:
The adaptive filter shown in Figure 7.1 updates the coefficients of the digital filter to optimize some
predetermined performance criterion. The most commonly used performance function is
based on the mean-square error (MSE).

5.10 Audio Processing:

The two principal human senses are vision and hearing. Correspondingly,much of DSP is related to image
and audio processing. People listen toboth music and speech. DSP has made revolutionary changes in
both these areas.

5.10.1 Music Sound processing:

The path leading from the musician's microphone to the audiophile's speaker is remarkably long. Digital
data representation is important to prevent the degradation commonly associated with analog storage and
manipulation. This is very familiar to anyone who has compared the musical quality of cassette tapes with
compact disks. In a typical scenario, a musical piece is recorded in a sound studio on multiple channels or
tracks. In some cases, this even involves recording individual instruments and singers separately. This is
done to give the sound engineer greater flexibility in creating the final product. The complex process of
combining the individual tracks into a final product is called mix down. DSP can provide several
important functions during mix down, including: filtering, signal addition and subtraction, signal editing,
etc. One of the most interesting DSP applications in music preparation is artificial reverberation. If
the individual channels are simply added together, the resulting piece sounds frail and diluted, much as if
the musicians were playing outdoors. This is because listeners are greatly influenced by the echo or
reverberation content of the music, which is usually minimized in the sound studio. DSP allows artificial
echoes and reverberation to be added during mix down to simulate various ideal listening environments.
Echoes with delays of a few hundred milliseconds give the impression of cathedral likelocations. Adding
echoes with delays of 10-20 milliseconds provide the perception of more modest size listening rooms.

5.10.2 Speech generation:

Speech generation and recognition are used to communicate between humans and machines. Rather than
using your hands and eyes, you use your mouth and ears. This is very convenient when your hands and
eyes should be doing something else, such as: driving a car, performing surgery, or (unfortunately) firing
your weapons at the enemy. Two approaches are used for computer generated speech: digital
recording and vocal tract simulation. In digital recording, the voice of a human speaker is digitized
and stored, usually in a compressed form. During playback, the stored data are uncompressed and
converted back into an analog signal. An entire hour of recorded speech requires only about three me

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

gabytes of storage, well within the capabilities of even small computer systems. This is the most common
method of digital speech generation used today. Vocal tract simulators are more complicated, trying to
mimic the physical mechanisms by which humans create speech. The human vocal tract is an acoustic
cavity with resonate frequencies determined by the size and shape of the chambers. Sound originates in
the vocal tract in one of two basic ways, called voiced and fricative sounds. With voiced sounds, vocal
cord vibration produces near periodic pulses of air into the vocal cavities. In comparison, fricative sounds
originate from the noisy air turbulence at narrow constrictions, such as the teeth and lips. Vocal tract
simulators operate by generating digital signals that resemble these two types of excitation. The
characteristics of the resonate chamber are simulated by passing the excitation signal through a digital
filter with similar resonances. This approach was used in one of the very early DSP success stories, the
Speak & Spell, a widely sold electronic learning aid for children.

5.10.3 Speech recognition:

The automated recognition of human speech is immensely more difficult than speech generation. Speech
recognition is a classic example of things that the human brain does well, but digital computers do poorly.
Digital computers can store and recall vast amounts of data, perform mathematical calculations at blazing
speeds, and do repetitive tasks without becoming bored or inefficient. Unfortunately, present day
computers perform very poorly when faced with raw sensory data. Teaching a computer to send you a
monthly electric bill is easy. Teaching the same computer to understand your voice is a major
undertaking. Digital Signal Processing generally approaches the problem of voice recognition in two
steps: feature extraction followed by feature matching. Each word in the incoming audio signal is
isolated and then analyzed to identify the type of excitation and resonate frequencies. These parameters
are then compared with previous examples of spoken words to identify the closest match. Often, these
systems are limited to only a few hundred words; can only accept speech with distinct pauses between
words; and must be retrained for each individual speaker. While this is adequate for many
commercialapplications, these limitations are humbling when compared to the abilities of human hearing.
There is a great deal of work to be done in this area, with tremendous financial rewards for those that
produce successful commercial products.

5.11 Image Enhancement:

5.11.1 Spatial domain methods:
Suppose we have a digital image which can be represented by a two dimensional random field A’y ^ .
t r
An image processing operator in the spatial domain may be expressed as a mathematical function L
applied to the image f to produce a new image -v^ “ ^ ^ 'X' -v^ - as follows.
g{x,y) = T\{x,y)_
The operator T applied on f ( x , y) may be defined over:
(i) A single pixel (x,y) . In this case T is a grey level transformation (or mapping) function.
(ii) Some neighbourhood of ( x , y).
(iii) T
may operate to a set of input images instead of a single image.
EXample 1
The result of the transformation shown in the figure below is to produce an image of higher contrast
than the original, by darkening the levels below m and brightening the levels above m in the original
image. This technique is known as contrast stretching.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Example 2
The result of the transformation shown in the figure below is to produce a binary image.
s = T(r)
Frequency domain methods
Let g( x, y) be a desired image formed by the convolution of an image f (x, y) and a linear, position
invariant operator h(x,y), that is:
g(x,y) = h(x,y)2f(x,y)
The following frequency relationship holds:
G(u, i’) = H (a, v)F(u, i’)
We can select H (u, v) so that the desired image
g(x,y) = 3 _ 1 i$(ii,v)F(u,v)
exhibits some highlighted features of f (x,y) . For instance, edges in f (x,y) can be accentuated by
using a function H(u,v) that emphasises the high frequency components of F(u,v) .

Glossary:

Sampling Rate:

The No. of samples per cycle given in the signal is termed as sampling rate of the signal .The samples
occur at T equal intervals of Time.

Sampling Theorem:

Sampling Theorem states that the no. of samples per cycle should be greater than or equal to twice that of
the frequency of the input message signal.

Sampling Rate Conversion:

The Sampling rate of the signal may be increased or decreased as per the requirement and application. This
is termed as sampling rate Conversion.

Visit : www.EasyEngineeering.net
 Visit : www.EasyEngineeering.net

Decimation:

The Decrease in the Sampling Rate are termed as decimation or Downsampling. The No. of Samples per
Cycle is reduced to M-1 no. of terms.

Interpolation:

The Increase in the Sampling rate is termed as Interpolation or Up sampling. The No. of Samples per
Cycle is increased to L-1 No. of terms.

Polyphase Implementation:

If the Length of the FIR Filter is reduced into a set of smaller filters of length k. Usual upsampling process
Inserts I-1 zeros between successive Values of x(n). If M Number of Inputs are there, Then only K
Number of Outputs are non-zero. These k Values are going to be stored in the FIR Filter.

Narrow band Filters:

If we want to design a narrow passband and a narrow transition band, then a lowpass linear phase FIR
filters are more efficiently implemented in a Multistage Decimator – Interpolator.

Visit : www.EasyEngineeering.net