1.

OBJECTIVES

Design the Linear Phase Finite Impulse Response (FIR) Filter and a lowpass filter and
bandpass filter through Window Design Technique of FIR.
2.

3.

COMPONENTS

Dekstop PC

Matlab 5.3 with Signal Processing Toolbox

THEORIES
3.1. LINEAR PHASE FIR FILTER
Among all the obvious advantages that digital filters offer, the FIR filter can guarantee
linear phase characteristics. There are many commercially available software packages
for filter design. However, without basic theoretical knowledge of the FIR filter, it will be
difficult to use them.
Filter coefficients:

Filter structure
A causal FIR filter whose impulse response is symmetrical is guaranteed to have a
linear phase response. (Even symmetry & Odd symmetry)

Frequency Response of an FIR Filter

To fully

design

and implement a filter five steps are required:

(1)

Filter specification.

(2)

Coefficient calculation.

(3)

Structure selection.

(4)

Simulation (optional).

(5)

Implementation.

There are several different methods available, the most popular are:
Window method.
Frequency sampling.
Parks-McClellan.
We will just consider the window method.
Linear phase is a one type of a filter. Filter need to modify a signal's magnitudespectrum
linear

when preserving the signal's time-domain waveform as much as possible. This

phase filter can be divided into four type of FIR:

symmetric sequence of odd length

symmetric sequence of even length

anti-symmetric sequence of odd length

anti-symmetric sequence of even length

There are four possible situation: filter length even or odd, and impulse response is
either

symmetric or antisymmetric :

FIGURE 2.0
3.2. FIR I AND FIR II TYPE
The symmetric coefficients shown that the frequency responses are of the
following form:
FIR I (M is even, sequence is symmetric and of odd length)

However, this system has linear phase (the quantity inside the parenthesis is a real
quantity) and the phase delay is M/2 samples.

 FIR II (M is odd, the sequence is symmetric and of even length)

Note
that this of
the

form

H() = ejG(), where =M/2, and G() is the real quantity

(the summation term) Output is delayed by M/2 samples
3.3. FIR III AND FIR IV
For anti symmetric sequences, we have h[n]=-h[M-n], which gives us sin terms is the
summation expression:
FIR III (M is even, the sequence is antisymmetric and of odd length)

FIR
IV (M is
odd, the
sequence is antisymmetric and of even length)

In
both
cases, the phase response is of the form ()=-(M/2) + /2, hence generalized
linear phase. Again, in all of theses cases, the filter output is delayed by M/2 samples.
Also, for all cases, if G()<0, an additional term is added to the phase, which causes
the samples to be flipped.
The symmetric coefficients shown that the frequency responses are of the
following form:
FIR I (M is even, sequence is symmetric and of odd length)

However, this system has linear phase (the quantity inside the parenthesis is a real
quantity) and the phase delay is M/2 samples.
FIR II (M is odd, the sequence is symmetric and of even length)

Note
that this of
the

form

H() = ejG(), where =M/2, and G() is the real quantity

(the summation term) Output is delayed by M/2 samples
The symmetric coefficients shown that the frequency responses are of the
following form:
FIR I (M is even, sequence is symmetric and of odd length)

However, this system has linear phase (the quantity inside the parenthesis is a real
quantity) and the phase delay is M/2 samples.
FIR II (M is odd, the sequence is symmetric and of even length)

Note
that this of
the

form

H() = e-jG(), where =M/2, and G() is the real quantity

(the summation term) Output is delayed by M/2 samples
The symmetric coefficients shown that the frequency responses are of the
following form:
FIR I (M is even, sequence is symmetric and of odd length)

However, this system has linear phase (the quantity inside the parenthesis is a real
quantity) and the phase delay is M/2 samples.
FIR II (M is odd, the sequence is symmetric and of even length)

Note
that this of
the

form

H() = ejG(), where =M/2, and G() is the real quantity

delayed by M/2 samples

4.

PROCEDURES

1) Open the MatLab software

2) Create the Matlab.m file (Refer to appendix)
3) Create the manipulated variables
4) Generate the output signal
5) Analyze and observe the output signal

(the summation term) Output is

5.

RESULTS

Part A
Properties of Linear-Phase Finite Impulse Response (FIR) Filters.
Type-1 FIR Filters
MATLAB script:
function [Hr,w,a,L] = Hr_Type1(h);
% Computes Amplitude response Hr(w) of a Type-1 LP FIR filter
% ----------------------------------------------------------% [Hr,w,a,L] = Hr_Type1(h)
% Hr = Amplitude Response
% w = 500 frequencies between [0 pi] over which Hr is computed
% a = Type-1 LP filter coefficients
% L = Order of Hr
% h = Type-1 LP filter impulse response
M = length(h); L = (M-1)/2;
a = [h(L+1) 2*h(L:-1:1)];
% 1x(L+1) row vector
n = [0:1:L];
% (L+1)x1 column vector
w = [0:1:500]'*pi/500;
Hr = cos(w*n)*a';

>>h = [-4 1 -1 -2 5 6 5 -2 -1 1 -4]

h=
-4

1 -1

-2

5 -2

-1

1 -4

>>M = length(h); n = 0:M-1

n =
0

-2

-8

>>[Hr,w,a,L] = Hr_Type1(h);
>>a,L
a =
6

10

-4

L =
5
>>amax = max(a)+1; amin = min(a)-1;
>>subplot(2,2,1); stem(n,h); axis([-1 2*L+1 amin amax])
>>xlabel('n'); ylabel('h(n)'); title('Impulse Response')
7
>>subplot(2,2,3); stem(0:L,a); axis([-1 2*L+1 amin amax])
>>xlabel('n'); ylabel('a(n)'); title('a(n) coefficients')
>>subplot(2,2,2); plot(w/pi,Hr);grid

10

>>xlabel('frequency in pi units'); ylabel('Hr')

>>title('Type-1 Amplitude Response')
>>subplot(2,2,4);
>>[Z,P,K]=tf2zp(1,a);
>>zplane(Z,P)

Output Simulation:

Impulse Response
10

h(n)

-5

10

10

a(n) coefficients
10

a(n)

-5

Type-2 FIR Filters

MATLAB script:
function [Hr,w,b,L] = Hr_Type2(h);
% Computes Amplitude response of a Type-2 LP FIR filter
% ----------------------------------------------------% [Hr,w,b,L] = Hr_Type2(h)
% Hr = Amplitude Response
% w = frequencies between [0 pi] over which Hr is computed
% b = Type-2 LP filter coefficients
% L = Order of Hr
% h = Type-2 LP impulse response
M = length(h); L = M/2;
b = 2*[h(L:-1:1)];
n = [1:1:L]; n = n-0.5;
w = [0:1:500]'*pi/500;
Hr = cos(w*n)*b';

>> h = [-4 1 -1 -2 5 6 6 5 -2 -1 1 -4]

h=
-4

1 -1

-2

5 -2

-1

-4

>> M = length(h); n = 0:M-1

n =
0

>>[Hr,w,b,L] = Hr_Type2(h);
>> b,L
b =
12

10

-4

-2

-8

L =
6
>> bmax = max(b)+1; bmin = min(b)-1;
>> subplot(2,2,1); stem(n,h); axis([-1 2*L+1 bmin bmax])
>> xlabel('n'); ylabel('h(n)'); title('Impulse Response')
>> subplot(2,2,3); stem(1:L,b); axis([-1 2*L+1 bmin bmax])
>> xlabel('n'); ylabel('b(n)'); title('b(n) coefficients')
>> subplot(2,2,2); plot(w/pi,Hr);grid
>> xlabel('frequency in pi units'); ylabel('Hr')
>> title('Type-1 Amplitude Response')
>> subplot(2,2,4);
>> [Z,P,K]=tf2zp(1,b);
>> zplane(Z,P)

10

10

11

Output Simulation:

Impulse Response
10

h(n)

-5

6
n

10

12

10

12

b(n) coefficients
10

b(n)

-5

11

6
n

Type-3 FIR Filters

MATLAB script:
function [Hr,w,c,L] = Hr_Type3(h);
% Computes Amplitude response Hr(w) of a Type-3 LP FIR filter
% ----------------------------------------------------------% [Hr,w,c,L] = Hr_Type3(h)
% Hr = Amplitude Response
% w = frequencies between [0 pi] over which Hr is computed
% c = Type-3 LP filter coefficients
% L = Order of Hr
% h = Type-3 LP impulse response
M = length(h); L = (M-1)/2;
c = [2*h(L+1:-1:1)];
n = [0:1:L];
w = [0:1:500]'*pi/500;
Hr = sin(w*n)*c';

>> h = [-4 1 -1 -2 5 0 -5 2 1 -1 4]
h=
-4

1 -1

-2

-5

-1

>> M = length(h); n = 0:M-1

n =
0

-2

-8

>>[Hr,w,c,L] = Hr_Type3(h);
>> c,L
c =
0

10

-4

L =
5

>> cmax = max(c)+1; cmin = min(c)-1;

>> subplot(2,2,1); stem(n,h); axis([-1 2*L+1 cmin cmax])
>> xlabel('n'); ylabel('h(n)'); title('Impulse Response')
>> subplot(2,2,3); stem(0:L,c); axis([-1 2*L+1 cmin cmax])
>> xlabel('n'); ylabel('c(n)'); title('c(n) coefficients')
>> subplot(2,2,2); plot(w/pi,Hr);grid
>> xlabel('frequency in pi units'); ylabel('Hr')
>> title('Type-1 Amplitude Response')
>> subplot(2,2,4);
>> [Z,P,K]=tf2zp(1,c);
>> zplane(Z,P)
12

10

Output Simulation:
Impulse Response

Type-1 Amplitude Response

25

10

20
15
10
Hr

h(n)

5
0

-5

-5
0

-10

10

0.1

0.2

0.3

0.4
0.5
0.6
frequency in pi units

0.7

0.8

0.5
Imaginary Part

c(n)

c(n) coefficients
10

-0.5

-5

-1
0

10

-1

13

-0.5

0
Real Part

0.5

0.9

Type-4 FIR Filters

MATLAB script:
function [Hr,w,d,L] = Hr_Type4(h);
% Computes Amplitude response of a Type-4 LP FIR filter
% ----------------------------------------------------% [Hr,w,d,L] = Hr_Type4(h)
% Hr = Amplitude Response
% w = frequencies between [0 pi] over which Hr is computed
% d = Type-4 LP filter coefficients
% L = Order of d
% h = Type-4 LP impulse response
M = length(h); L = M/2;
d = 2*[h(L:-1:1)];
n = [1:1:L]; n = n-0.5;
w = [0:1:500]'*pi/500;
Hr = sin(w*n)*d';

>> h = [-4 1 -1 -2 5 6 -6 -5 2 1 -1 4]
h=
-4

-1 -2

6 -6

-5

-1

>> M = length(h); n = 0:M-1

n =
0

>>[Hr,w,d,L] = Hr_Type4(h);
>> d,L
d =
12

10

-4

-2

-8

L =
6
>> dmax = max(d)+1; dmin = min(d)-1;
>> subplot(2,2,1); stem(n,h); axis([-1 2*L+1 dmin dmax])
>> xlabel('n'); ylabel('h(n)'); title('Impulse Response')
>> subplot(2,2,3); stem(1:L,d); axis([-1 2*L+1 dmin dmax])
>> xlabel('n'); ylabel('d(n)'); title('d(n) coefficients')
>> subplot(2,2,2); plot(w/pi,Hr);grid
>> xlabel('frequency in pi units'); ylabel('Hr')
>> title('Type-1 Amplitude Response')
>> subplot(2,2,4);
>> [Z,P,K]=tf2zp(1,d);
14
>> zplane(Z,P)

10

11

Impulse Response
10

h(n)

-5

6
n

10

12

10

12

Output Simulation:

d(n) coefficients
10

d(n)

-5

4
15

6
n

Part B
Designing the Finite Impulse Filter (FIR) using Window Techniques.
MATLAB script:
1.

Function of ideal lowpass impulse response:

function hd=ideal_lp(wc,M);
% Ideal LowPass filter computation
% -------------------------------% [hd] = ideal_lp(wc,M);
% hd = ideal impulse response between 0 to M-1
% wc = cutoff frequency in radians
% M = length of the ideal filter
alpha = (M-1)/2;
n = [0:1:(M-1)];
m = n - alpha +eps;
% add smallest number to avoid divided by zero
hd = sin(wc*m)./(pi*m);

2.

Function of frequency magnitude (dB):

function [db,mag,pha,grd,w] = freqz_m(b,a)

% Modified version of freqz subroutine
% -----------------------------------% [db,mag,pha,grd,w] = freqz_m(b,a)
% db = relative magnitude in dB computed over 0 to pi radians
% mag = absolute magnitude computed over 0 to pi radians
% pha = Phase response in radians over 0 to pi radians
% grd = Group delay over 0 to pi radians
% w = 501 frequency samples between 0 to pi radians
% b = numerator polynomial of H(z) (for FIR: b=h)
% a = denominator polynomial of H(z) (for FIR: a=[1])
[H,w] = freqz(b,a,1000,'whole') ;
H = (H(1:1:501))';
w = (w(1:1:501))';
mag = abs(H);
db = 20*log10((mag+eps)/max(mag));
pha = angle(H);
grd = grpdelay(b,a,w);

16

3.

Desired digital bandpass filter as required:

% Design using the given digital passband filter

ws1 = 0.2*pi; wp1 = 0.35*pi; wp2 = 0.65*pi; ws2 = 0.8*pi; As = 60;
tr_width = min((wp1-ws1),(ws2-wp2));
M = ceil(11*pi/tr_width) +1
n = [0:1:M-1];
wc1 = (ws1+wp1)/2; wc2 = (wp2+ws2)/2;
hd = ideal_lp(wc2,M) - ideal_lp(wc1,M);
w_bla = (blackman(M))';
h = hd.*w_bla;
[db,mag,pha,grd,w] = freqz_m(h,[1]);
delta_w =2*pi/1000;
Rp = -min(db(wp1/delta_w+1:1:wp2/delta_w)) % Actual Passband Ripple
As = -round(max(db(ws2/delta_w+1:1:501))) % Min Stopband Attenuation

Output Command from MATLAB:

M =
75
Rp =
0.0030
As =
75

17

Ideal Impulse Response

0.4
0.3

hd(n)

0.2
0.1
0
-0.1
-0.2
-0.3
-0.4

10

20

30

40

50

60

70

60

70

Actual Impulse Response

Output Simulation:

0.4
0.3

h(n)

0.2
0.1
0
-0.1
-0.2
-0.3
-0.4

10

20

30
18

40

50

Designing the ideal digital differentiator using function Hr_Type3(h) file.

Hr_Type3(h) that has been created from the first part of lab:
function [Hr,w,c,L] = Hr_Type3(h);
% Computes Amplitude response Hr(w) of a Type-3 LP FIR filter
% ----------------------------------------------------------% [Hr,w,c,L] = Hr_Type3(h)
% Hr = Amplitude Response
% w = frequencies between [0 pi] over which Hr is computed
% c = Type-3 LP filter coefficients
% L = Order of Hr
% h = Type-3 LP impulse response
M = length(h); L = (M-1)/2;
c = [2*h(L+1:-1:1)];
n = [0:1:L];
w = [0:1:500]'*pi/500;
Hr = sin(w*n)*c';

Desired filter to be simulated:

>> M = 21; alpha = (M-1)/2; n = 0:M-1

n =
Columns 1 through 16
0

12

13

14

15

10

11

Columns 17 through 21
16

17

18

19

20

>> hd = (cos(pi*(n-alpha)))./(n-alpha); hd(alpha+1)=0;

>> w_ham = (hamming(M))'; h = hd .* w_ham; [Hr,w,P,L] = Hr_Type3(h);
>> subplot(2,2,1); stem(n,hd); title('Ideal Impulse Response')
>> axis([-1 M -1.2 1.2]); xlabel('n'); ylabel('hd(n)')
>> subplot(2,2,2); stem(n,w_ham);title('Hamming Window')
>> axis([-1 M 0 1.2]); xlabel('n'); ylabel('w(n)')
>> subplot(2,2,3); stem(n,h);title('Actual Impulse Response')
>> axis([-1 M -1.2 1.2]); xlabel('n'); ylabel('h(n)')
>> subplot(2,2,4);plot(w/pi,Hr/pi); 19
title('Amplitude Response');grid;
>> xlabel('frequency in pi units'); ylabel('slope in pi units'); axis([0
1 0 1]);

Ideal Impulse Response

1

hd(n)

0.5
0
-0.5
-1
0

10
n

12

14

16

18

20

16

18

20

Actual Impulse Response

1
0.5

h(n)

Output Simulation:

0
-0.5
-1
0

6
20

10
n

12

14

6.

DISCUSSIONS
6.1. PART 1 (TYPE 1 - TYPE 4 LINEAR FIR FILTER)
6.1.1.

MATLAB COMMANDS

In order to get the wanted output, some Matlab command which corresponding the FIR equation
are entered into Matlab to be processed. There are including some constants and variables.
Below are constants which been used in Part 1 demonstrations:
L is order of Hr. L is written as, L = (M-1)/2.
w is frequencies between [0, pi] over which Hr is computed. w is w = [0:1:500] *pi/500.
n written as n = [0:1:L]

Type-1 FIR Filters

For type Hr, equation used is, Hr = cos (w*n) * a, where, a = [h(L+1) 2*h(L:-1:1)].

21

Type-2 FIR Filters

For type Hr, equation used is, Hr = cos (w*n) * b, where, b = 2 * [h(L:-1:1)].

Type-3 FIR Filters

For type Hr, equation used is, Hr = sin (w*n) * c, where, c = [ 2 * h (L+1: -1 : 1) ].

Type-4 FIR Filters

For type Hr, equation used is, Hr = sin (w*n) * d, where, d = 2 * [h(L: -1:1)].

6.1.2

OUTPUT GRAPHS

From graph generated by Matlab, there are no restrictions on Hr(w) either w=0 or w=pi. System
poles

shows that there are three (3) poles on the right of real part, one x plane while the other

two symmetrically near y plane. On the negative side of real part, there are two more
on the x-plane.

Since poles present on the positive side of real part, system is not stable.

cannot analyze zeros

poles
We

because error in coding. It shows that we were missing library to

generate zero coding.

From the plots, Hr(w) is zero at w = pi. Poles coordination is same as FIR Filters type-1 except
that on the negative side of real part, instead of the poles placed on x-plane, poles floating up to
0.5 and to -0.5. System also unstable.
From the plots, Hr(w) = 0 at w = 0 and w = pi. Poles coordination are same with FIR Filters
Type-3.
From plots, it can be observed that Hr(w) is zero at w = 0. The poles pattern are same with
Type-2 and

Type-3.

6.2. PART 2
For the second part of the lab, we were asked to construct a digital bandpass FIR filter using the

22

following specifications;

lowerstopband
edge:

1s 0.2

As 60dB

lowerpassband
edge:

1p 0.35

Rp 1dB

upperpassband
edge:

2p 0.65

Rp 1dB

upperstopband
edge:

2s 0.8

As 60dB

Below is the image of the digital bandpass filter that must be produced. Notice that the highlighted
ones are the selected bandpass filter based on the requirements given above.

To do this, we need to identify which 2

1
1
2p
2s
1
2 1
window design to be used. First, we need to measure the
bands, and

bandwidth of the bands. The two transition

must be the same in the window design (there is no independent control over and . For

, we can use the Blackman window.

1 2

Blackman window design uses the same function of Hann window and Hamming window, except it
contains a second harmonic term given as follows;

23

 2n
4n
0.42 0.5cos
0.08cos

n
M 1
M 1

24

;0 n M 1
;otherwise

We also need the ideal bandpass filter hd n impulse response function, . Therefore, the
MATLAB

routine ideal_lp(wc,M) is sufficient to determine the impulse response of an ideal

bandpass filter.

We also include the modified function of the frequency-domain plots or freqz_m,

which returns the

magnitude response in absolute as well as in relative dB scale, the phase response and

the group delay

response.

Based on the requirements given, using the command window of MATLAB, the following string of
code are entered so that it declare the position or the exact bandwidth of the bandpass filter:

ws1 = 0.2*pi; wp1 = 0.35*pi; wp2 = 0.65*pi; ws2 = 0.8*pi; As = 60;

Next, the transition width is calculated by p

s finding the minimum difference between and ;
tr_width = min((wp1-ws1),(ws2-wp2));

Summary of commonly used window function characteristics.

Window
Name
Rectangular
Bartlett
Hanning
Hamming
Blackman

Transition

Approximate
4
M
8
M
8
M
8
M
12
M

Width,
Exact Values
1.8
M
6.1
M
6.2
M
6.6
M
11
M

Min. Stopband
Attenuation
21 dB
25 dB
44 dB
53 dB
74 dB

Next, find the value of filter length, M for the transitional width, (this is decided from the
summary table given above using the exact values),

M = ceil(11*pi/tr_width) +1

25

which returns the following results from the MATLAB output command;
M =
75

Next, entered the center frequency of the bandpass filter given as follows;

n = [0:1:M-1];
wc1 = (ws1+wp1)/2; wc2 = (wp2+ws2)/2;

Then, we entered the ideal bandpass filter function. Since there are two different center frequency, the
value of hd is calculated as follows;

hd = ideal_lp(wc2,M) - ideal_lp(wc1,M);

The values of hd will stored the attached value of number of pulses, n given above.
Next, run the Blackman window as blackman(M) and multiplied with the given hd, and stored in
h.

w_bla = (blackman(M))';
h = hd.*w_bla;

The function of freqz_m is used in this time with the stored value of the previous h.

[db,mag,pha,grd,w] = freqz_m(h,[1]);
is given by the default formula as ;

2
1000

delta_w =2*pi/1000;

26

Then, finally we obtained the actual bandpass ripple, Rp and minimum stopband attenuation, As. These
are given by the formula as follows;

Rp 20log10

1 1
0
1 1

As 20log10

2
0
1 1

Rp = -min(db(wp1/delta_w+1:1:wp2/delta_w))
As = -round(max(db(ws2/delta_w+1:1:501)))

which comes the results as follows;

Rp =
0.0030

As =
75

Then, to construct the following output simulation, the following string code of MATLAB command
is inputted as follows;

subplot(2,2,1); stem(n,hd); title('Ideal Impulse Response')

axis([0 M-1 -0.4 0.5]); xlabel('n'); ylabel('hd(n)')
subplot(2,2,2); stem(n,w_bla); title('Blackman Window')
axis([0 M-1 0 1.1]); xlabel('n'); ylabel('w(n)')
subplot(2,2,3); stem(n,h); title('Actual Impulse Response')
axis([0 M-1 -0.4 0.5]); xlabel('n'); ylabel('h(n)')
subplot(2,2,4); plot(w/pi,db); axis([0 1 -150 10]);

27

title('Magnitude Response in db');grid;

xlabel('frequency in pi units'); ylabel('Decibels')

The output of this simulation are shown in the results section.

HAMMING WINDOW
Hr is amplitude response, equation used is, Hr = sin (w*n)*c. Where,
L is order of Hr. L is written as, L = (M-1)/2 and M = length (h).
w is frequencies between [0, pi] over which Hr is computed. w is w = [0:1:500]

*pi/500.

n written as n = [0:1:L]
c is type-3 low pass filter coefficients. c = [2*h(L+1:-1:1]

Frequency response of Hamming Window is sinusoidal in shape. This because it defines by w(n) =
0.54 - 0.46 cos (2pin/N). For n = 1, 2, ..., N - 1. In time domain, the Hamming window does not get as
close to zero near the edges.

7.

CONCLUSIONS
FIR filters is useful filter used in digital application due to some of its advantages. To
analyze this filter, we can use Matlab to analyze it.
To design FIR filter, user need to know the filter specification to be simulated. Then to
know the coefficient calculation followed by structure selection. Simulation that have

28

been done in this laboratory can be considered as important step. In this laboratory, the
signal have been designed by using Matlab. Finally is the implementation of the signal.
There are four type of FIR as been discussed in result part and discussion. Type 1,
symmetric sequence of odd length, Type 2, symmetric sequence of even length, Type 3,
anti-symmetric sequence of odd length and Type 4,anti-symmetric sequence of even
length.
With Matlab also we can construct bandpass filter, in this lab two windows model have
been used, Blackman and Hammings. There are other windows model can be used
although not demonstrated and discussed in this lab. There are Rectangular, Bartlett
and

Hanning windows. All with their special characteristics can be used in certain

application and function.

8.

REFERRENCES
Vinay K. I. and John J. P., Digital Signal Processing Using Matlab (3 rd ed.), Cengage
Learning, US
J. W. Cooley and J. W. Tukey. An algorithm for the machine computation of

complex

29

Fourier series. Mathematical Computations, 19:297301, April 1965.

R. E. Crochiere and L. R. Rabiner. Multirate Digital Signal Processing. Prentice Hall,
Englewood Clis, NJ, 1983.
C. de Boor. A Practical Guide to Splines. Springer-Verlag, 1978.
J. L. Flanagan et al. Speech coding. IEEE TransactionsonCommunications,
COM27:710736, April 1979.

30