Speech signal analysis for linear filter banks of different orders

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
_______________________________________________________________________________________
Volume: 04 Issue: 06 | June-2015, Available @ http://www.ijret.org 248
SPEECH SIGNAL ANALYSIS FOR LINEAR FILTER BANKS OF
DIFFERENT ORDERS
Somashekhar Shankarayya Pujar1
, Shridhar S. Kuntoji2
, Mahiboob3
1
Department of Electronics and Communication Engineering, Basaveshwara Engineering College Vidyageri,
Bagalkot-587103, India
2
3
Abstract
In speech signal processing using of filter banks is very important. The critical requirement is the sum of all the frequency
responses of the band-pass filters of the filter bank i.e. composite frequency response be flat with linear phase. This paper deals
with design and implementation of linear-phase FIR digital filters based filter bank flat with flat composite frequency response.
The design is based on special properties of FIR filters by which excellent frequency response can be achieved.
Keyword: Speech signal, FIR Filter, Composite frequency response.
-------------------------------------------------------------------***-------------------------------------------------------------------
1. INTRODUCTION
In a variety of speech processing systems filter banks are
used to perform short time spectrum analysis. To cover a
desired portion of the speech band a set of band-pass filters
is designed. By simply adding together the outputs of the
band-pass filters, it is possible to approximate the spectrum
of the input speech signal.
It is possible to characterize the behavior of filter banks by
considering the composite frequency response. Since,
ideally, the combined output should be equal to the input,
then we desire that the composite frequency response have
constant magnitude and linear phase over the desired band
of frequencies. Earlier work emphasizes the relative phases
between channels which is important in achieving a flat
composite frequency response. Using this method excellent
overall response can be obtained for both IIR and FIR
digital filters in filter banks in which the center frequencies
are uniformly spaced. This method is not easily extendable
to non-uniformly spaced filter banks.
2. SPEECH ANALYSIS
Speech signals are basically partitioned into voiced
speech segments and unvoiced speech segment . A voiced
speech segment is also known as pitch of voiced speech. It
has high energy content and is periodic in nature. The
unvoiced part of the speech looks like a random noise
with no periodicity. Some parts of the speech that are
neither voiced nor unvoiced are called transition segments.
The speech can be analysed either by Time domain method
and Frequency domain method.
3. EXAMPLE OF BPF DESIGN:
Hd(w)=
Cuttoff frequency of 1 = , 2 = rad/samples. Length
of the filter M = 25
hd(n) = ejwn
d
hd(n) = [ ejwn
e-jwT
d +
ejwn
e-jwT
d ]
= [ ejw ( n - T)
d +
ejw ( n – T )
d ]
= [ +
[
hd(n) = for n ≠ T
when n = T equation will becomes as
hd(n) = [ d + d ]
hd(n) =
hd(w)=

_______________________________________________________________________________________
T= where M=25
T=
T=12
hd(w)=
n=1,2,3,4,5,..................25.
Result
Fig: 3. Band Pass Filter Output
4. FILTER BANK DESIGN
To design a filter bank using FIR filters, we must first
determine the range of frequencies to be covered by the
composite response. Let us assume that these are denoted
ωmin and ωmax, where ωmax ≤ . Now, if there are a
total of Nf, f ilters, we must choose the bandwidths and
center frequencies so that the entire range of frequencies
ωmin ωmax is covered. This is depicted in Fig 2.3.1
for the case Nf = 3. This figure shows the ideal responses
for each bandpass filter;
Fig:4 A typical nonuniform filter bank.
4.1 Design Examples
In this section, we illustrate the use of the principles with
examples of both uniform and non-uniform filter banks. For
all the examples, the sampling rate is assumed to be 8
kHz.
Example 4.1
Suppose that we wish to design a bank of 15 equally
spaced filters that covers the range 200 to 3200Hz. Then we
find that the cutoff frequency for all the low-pass filters
is
fo = = 100Hz
The center frequencies are
fck = = 100(2k+1)Hz k=1,2,3………15.
Table: 4.1 Values of center frequencies
K 8 0.53125
1 0.09375 9 0.59370
2 0.15625 10 0.65625
3 0.21875 11 0.71875
4 0.28125 12 0.78125
5 0.34375 13 0.84375
6 0.40625 14 0.90625
7 0.46875 15 0.96875
4.1.1 The specifications are,
1. Attenuation= 60dB.
2. Cutoff frequency= 100Hz.
3. The widest transition band that is reasonable is 200
Hz.
4. Order must N= 150.
5. Adjacent channels cross at 0.5 amplitude.
6. Filters will merge together.
7. The deviation is from unity.

_______________________________________________________________________________________
Fig: 4.1.1 FIR Filter Banks
Fig: 4.1.2 Composite Frequency Response
Figure 4.1.2 shows the composite response of the filter
bank. It is clear that the filters merge together very well at
the edges of the frequency bands. Indeed, the deviation
from unity is less than or equal to the peak approximation
error, =0.001.
Example 4.2
A non-uniform spacing of the filters is often used to
exploit the ear's decreasing frequency resolution with
increasing frequency. Suppose that we wish to cover the
same range 200 to 3200 Hz as in Example 4.1, but we
wish to use only four octave band filters. That is, each
successive filter will have a bandwidth twice the
bandwidth of the previous filter. We find that the lowest
frequency channel has cutoff frequency.
fo= = = 100Hz
In general, the cutoff frequencies of the prototype low-pass
filters are
fpk= = fo k= 1, 2, 3, 4
Table: 4.2 Values of center frequencies
k
1 0.03125
2 0.0625
3 0.125
4 0.25
4.2.1 The specifications are,
1. No. of filters=4.
2. Spacing = Twice the bandwidth of the previous
filter.
3. Coverage range = 200 to 3200 Hz.
4. Order N= 150.
Fig: 4.2.1, 4 Non-uniform FIR Filter Banks for N = 150
Fig: 4.2.2 Composite Frequency Response
Example 4.3
Suppose that all the parameters remain the same as in
Example 4.2 except that we require narrower transition
regions. This means that a larger value of N is required.
Figure III.4.1 shows the filter bands corresponding to the
parameters of Example 4.2 except that N =301 and =
0.012082 (transition width is 116 Hz). The sharper
transitions are apparent in Fig. 4.3.1 shows that the
composite response

_______________________________________________________________________________________
Fig: 4.3.1, Four Non-uniform FIR Filter Banks for N = 301
and Composite frequency.
Example 4.4
We have assumed throughout that the transition width was
less than twice the smallest lowpa -ss cutoff frequency. In
our examples, this constraint required that N be at least 150.
In this case, all the parameters were the same as in
Examples 2.2 and 2.3, except in the case of N=101and
=0.0362465. The transition width is 348 Hz, which is
much greater than twice the cutoff frequency of the first
low-pass filter. It is clear that reasonable filters are
obtained for the wider bandwidth filters however, the lowest
filter does not attain unity response anywhere in its pass-
band. The preceding examples make it abundantly clear
that, for sufficiently long impulse responses, the composite
filter bank response can be very flat.
Fig: 4.4, Four Non-uniform FIR Filter Banks for N = 101
and Composite frequency.
Where design techniques for IIR filter banks were
discussed, the best results achieved for the composite
response were approximately 1-dB peak-to-peak ripple for
uniform bandwidths and about 2.5-dB peak-to-peak ripple
for non-uniform bandwidths. This is in contrast to the
results of the examples of this section, where the peak-to-
peak ripple in the composite response was about 0.0274 dB
for all the filter banks independent of how the bandwidths
were chosen. This, together with the precise linear phase
that is easily achieved, makes the FIR filter banks superior
to what can be achieved for IIR filter banks. The price that
is paid for this is that rather large values of N are
required to achieve sharp transitions. However, the values
of N used in the previous examples are certainly not
unreasonable if the filters are implemented by FFT
convolution methods or in special purpose hardware.
5. RESULTS AND DISCUSSION
In this paper totally four filter banks were designed using
FIR filter. One is linearly spaced and remaining three are
non-linearly spaced. These filter banks are tested with some
known signals and monotone signals, such as cosine signals
with certain value of frequencies and also vowels.
When 1000Hz cosine signal is applied to all the filter banks.
The output is obtained in only one filter bank which is
having the bandwidth in the range of 1000Hz (860-1243
Hz). The output can be found by observing the output
signals of all the FIR filter banks. The signal which is
having more periodic as compared to other signals will be
our required output. And in other way is by calculating the
energy of all the filter outputs. The filter output which is
having the higher energy is our required output. Later
monotone signals are applied to filters such as vowels (a, e,
i, o, u) and calculated the bandwidths of those signals.
5.1 Calculating the Bandwidths of monotone and
vowels signals.
Table 5.1.1 Result obtained when the input to the filter bank
(Linearly spaced filters) is sine signal of different
frequencies.
Sl/No Input
Frequency Hz
Bands Energy
1 100 37- 420 0.00028855
2 500 457-841 0.00022262
3 1000 660-1006 0.00015226
4 2000 1847-2213 0.0030
5 3000 2871-3190 0.0055
Table 5.1.2 Result obtained when isolated vowels are input
to the filter bank.
Sl/No Vowels of speaker 1 Bands Energy
1 A 256-659 662.0616
2 E 37-420 43204873
3 I 256-659 950.5838
4 O 37-420 2456.2
5 U 37-420 69.7421

_______________________________________________________________________________________
to the filter bank.
Sl/No Vowels of
speaker 2
Bands Energy
1 A 660-1006 61.6245
2 E 37-420 80.2513
3 I 37-420 99.8089
4 O 457-841 130.80
5 U 37-420 30.0643
to the filter bank.
Sl/No Vowels of
speaker 3
Bands Energy
1 A 731-1646 104.2574
2 E 731-1646 96.9293
3 I 731-1646 175.4988
4 O 330-932 486.1503
5 U 55-495 26.9864
5.2 Listening sounds of all the FIR filters of the filter bank
when isolated vowels are input it is observed that, Only
filters of lower bands are able to produce the given clear
speech stimuli as output. It means the output speech is same
as the input speech. Whereas in the higher bands the output
does not convey the actual speech. This is true for all higher
band filters. It is because vowels are low pass signals.
6. CONCLUSION
We have discussed a design method for filter banks
composed of linear digital filters. The method exploits the
linear-phase properties obtainable for such filters, as well
as the symmetry of the transition region that results from
the windowing method of design. We summarized this
method of design for the Kaiser window and illustrated the
filter bank design method with several examples. These
examples show the proposed design method has a great
deal of flexibility and that excellent response
characteristics can be achieved.
ACKNOWLEDGMENTS
This paper is made possible through the help and support
from the guide Dr. K. Sridhar for his guidance,
understanding, and patience during my M.Tech course at
B.E.C Bagalkot and during the all phases of this paper.
REFERENCES
[1] J. L. Flanagan, Speech Analysis, Synthesis and
Perception, Second Ed., New York: Springer-
Verlag, 1972.
[2] J. L. Flanagan and R. M. Golden, "Phase Vocoder,"
B.S.T.J., 45, No.9 (November 1966), pp. 1493-
1509.
[3] R. W. Schafer and L. R. Rabiner, "Design and
Simulation of a Speech Analysis Synthesis System
Based on Short Time Fourier Analysis," IEEE Trans.
Audio and Electroacoustics, AU-21, No.3 (June
1973), pp. 165-174.
[4] J. L. Flanagan and R. C. Lummis, "Signal Processing
to Reduce Multipath Distortion in Small Rooms," J.
Acoust. Soc. Am., 47, part 1, June 1970, pp. 1475-
1481.
[5] R. W. Schafer and L. R. Rabiner, "Design of Digital
Filter Banks for Speech Analysis," B.S.T.J., 50, No.
10 (December 1971) pp. 3097-3115.
[6] J. F. Kaiser, "Digital Filters," System Analysis by
bigital Computer, Ch. 7, F. F. Kuo and J. F. Kaiser,
eds., New York: John Wiley, 1966, pp. 228-243.
[7] J. F. Kaiser, "Nonrecursive Digital Filter Design
using the Is-Sinh Window Function," Proc. 1974
IEEE Int. Symp. on Cir. and Syst., San Francisco,
1974, pp. 20--23.

Speech signal analysis for linear filter banks of different orders

More Related Content

What's hot (18)

Viewers also liked (16)

Similar to Speech signal analysis for linear filter banks of different orders (20)

More from eSAT Journals (20)

Recently uploaded (20)

Speech signal analysis for linear filter banks of different orders