********************************color.m*******************************

function [cmin,cmax] = color(specgram);

v = 20*log10(abs(specgram));

hist(v);

[g,h] = hist(v);

i = max(g(:)); %the maximum number in histogram

p = median(h); %position of the maximum number

cmax = p + 15;

cmin = p – 15;

********************************spec.m*******************************

%to produce a spectrogram of signals
function Y = spec(a);
%a = normal(a);%a = silence(a);a = a(:,1);
[s,f,t]=specgram(a);

[cmin,cmax] = color(s);clim = [cmin cmax];

Y = imagesc(t,f,20*log10(abs(s)),clim), axis xycolormap(1-gray)

********************************main_spec.m*******************************

[ref,fs] = wavread(‘Menu.wav’);

[test,fs] = wavread(‘T_Menu.wav’);

[r,t] = dtw(ref,test,fs);

r = r(1:14000,1);

t = t(1:14000,1);

figure(1);

r1 = spec(r);

figure(2);

t1 = spec(t);

********************************************************************

******************************************************************

%wavlope.m

%%from try_we3.m

function envelope = wavlope(wavefile, Fs);

% Step 1: Load speech waveform

%[y,rate]=wavread(‘Menu.wav’);

y = wavefile;

rate = Fs;

y = y(:,1);

y = normal(y);

y = silence(y);

% Step 2:

% Generate lowpass filter coefficients for envelope extraction

% 2nd order elliptical filter, 30 Hz cutoff, 0.5 dB ripple,

% min stopband attuation 72 dB

Rp=0.5;

Rs=72;

fc=30/(rate/2);

[blp,alp]=ellip(2,Rp,Rs,fc);

% Step 3: half-wave rectify the waveform

y(find(y<0))=0;

y = normal(y);

% extract envelope by half-wave rectification & LP filtering

envelope=filter(blp, alp, y);

envelope = normal(envelope);

******************************************************************

%main_wavelope.m

[ref,fs] = wavread(‘Nadzirah.wav’);

[test,fs] = wavread(‘T_Nadzirah.wav’);

y = ref(:,1);

y = normal(y);

y = silence(y);

x = test(:,1);

x = normal(x);

x = silence(x);

we = wavlope(y,fs); %ref signal

we2 = wavlope(x,fs); %test signal

[d1,d2] = dtw(we,we2,fs);

%d1 is the original signal of ref signal

%d2 is the dtwed signal of test signal

figure(10);

%subplot(211);

plot(we);

hold on

plot(we2,’k');

title(‘Before DTW’);

legend(‘Ref Signal’,'Test Signal’)

figure(101);

%subplot(212);

plot(d1);

hold on

plot(d2,’k');

axis([ 0 20000 0 1]);

title(‘After DTW’);

legend(‘Ref Signal’,'Test Signal’)

%plot(we);

%hold on;

%plot(-envelope);

figure(1);

subplot(211);

plot(y);

title(‘Original wave signal’);

ylabel(‘Amplitude’);

subplot(212);

plot(we);

title(‘Waveform envelope of wave signal’);

ylabel(‘Amplitude’);

xlabel(‘Time,s’);

Figure 5: (a) Plot of the original wave signal

(b) Plot of the waveform envelope of the wave signal

Figure 6: (a) Plot of the original wave signal

(b) Plot of the waveform envelope of the wave signal

[a,Fs] = wavread(‘Anne.wav’);

a = normal(a);

A = a(8000:end,1);

A = silence(A);

%to obtain the fourier transform

ffta = fft(A,512);

w = (0:255)/256*(Fs/2);

N = 310;

W = 1000/(Fs/2);

%design a high pass filter

b = fir1(N,W,’high’);

[H,ww] = freqz(b,512);

figure(1);

plot(ww*Fs/(2*pi),abs(H));

%to filter noise

y = filter(b,1,A);

ffth = fft(y,512);

figure(2);

subplot(311);

plot(A);

subplot(312);

plot(w,abs(ffta(1:256)’)); %fourier transform%

subplot(313);

plot(w,abs(ffth(1:256)));

Figure 1 : High Pass filter

Figure 2 (a) : Plot of the original wave

Figure 2(b): Plot of the fourier transform of the wave

Figure 2(c): Plot of the fourier transform of the filtered wave

In voice recording, there will always be silence moment in the beginning and ending of the speech. This kind of silence cannot be avoided as it is normal for a person to take few seconds to speak after the button ‘Record’ is hit and also before the button ‘Stop’ is hit. This silence may also be known as the background noise as we actually recording the noise of the background while we’re in silence. To remove these silence, this approach is used instead of interactive filter.

%%%%% to remove silence in the beginning and in the end%%%%%

function [c,d] = first(sound);

%w is a vector of the wave file

w=sound;

i=1;

while w(i)<=0

i=i+1;

end

a=i;

b=w(i:length(w)); %filtered wave of beginning silence only

j=length(b);

while b(j)<=0

j=j-1;

end

c=j;

d=b(1:j); %filtered wave of beginning & ending silence

%%%%% To implement the function ‘first’ %%%%%%

function A = silence(sound);

B = double(sound);

C = normal(B);

%C = b(:,1);

[NX,NY] = size(C);

for i = 1:NX

for j = 1:NY

if  C(i)  <  0.2  &  C(i) >  -0.2

C(i) = C(i)*0;

end

end

end

[D,A] = first(C);

DTW also allows us to separate speech from silence in the input signal. The silence in the beginning and at the end of the speech especially cannot be avoided. There will be a moment of silence within those periods of recording. Advanced versions of DTW were developed to match sequences of words in connected speech which has been superceded by the HMM approach but in fact the search for interpretations in an HMM is a very similar process to the DP match used here. (from DTW4)

Example of DTW coding:

****************************************************************

%dtw.m

function [d1,d2x] = dtw(ref,test,sr);

% Load two speech waveforms of the same utterance (from TIMIT)

%[ref,sr] = wavread(r);

%[test,sr] = wavread(t);

ref = silence(ref);

test = silence(test);

%whos

d1 = ref(:,1);

d2 = test(:,1);

% Listen to them together:

ml = min(length(d1),length(d2));

% soundsc(d1(1:ml)+d2(1:ml),sr)

% or, in stereo

%soundsc([d1(1:ml),d2(1:ml)],sr)

% Calculate STFT features for both sounds (25% window overlap)

D1 = specgram(d1,512,sr,512,384);

D2 = specgram(d2,512,sr,512,384);

% Construct the ‘local match’ scores matrix as the cosine distance

% between the STFT magnitudes

SM = simmx(abs(D1),abs(D2));

% Look at it:

figure(100);

subplot(121)

imagesc(SM);

colormap(1-gray);

% You can see a dark stripe (high similarity values) approximately

% down the leading diagonal.

% Use dynamic programming to find the lowest-cost path between the

% opposite corners of the cost matrix

% Note that we use 1-SM because dp will find the *lowest* total cost

[p,q,C] = dp(1-SM);

% Overlay the path on the local similarity matrix

hold on; plot(q,p,’r'); hold off

% Path visibly follows the dark stripe

% Plot the minimum-cost-to-this point matrix too

subplot(122)

imagesc(C);

hold on; plot(q,p,’r'); hold off

% Bottom right corner of C gives cost of minimum-cost alignment of the two

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%C(size(C,1),size(C,2));

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%ans = 128.2873;

% This is the value we would compare between different

% templates if we were doing classification.

% Calculate the frames in D2 that are indicated to match each frame

% in D1, so we can resynthesize a warped, aligned version

D2i1 = zeros(1, size(D1,2));

for i = 1:length(D2i1); D2i1(i) = q(min(find(p >= i))); end

% Phase-vocoder interpolate D2′s STFT under the time warp

D2x = pvsample(D2, D2i1-1, 128);

% Invert it back to time domain

d2x = istft(D2x, 512, 512, 128);

% Listen to the results

% Warped version alone

%%%%%%soundsc(d2x,sr)

% Warped version added to original target (have to fine-tune length)

d2x = resize(d2x’, length(d1),1);

%%%%%%%soundsc(d1+d2x,sr)

% .. and in stereo

%%%%%%%soundsc([d1,d2x],sr)

% Compare to unwarped pair:

%%%%%%%soundsc([d1(1:ml),d2(1:ml)],sr)

%y(1025:length(x)) = 0;            % Zeroing out the higher coefficients

%xHat = ifwht(y);                  % Signal reconstruction using inverse WHT

%figure(‘color’,'white’);

%plot(d2x);

%hold on

%plot(d1,’r');

%xlabel(‘Sample index’);

%ylabel(‘ECG signal amplitude’);

%legend(‘Original Signal’,'Reconstructed Signal’);

***************************************************************

%main_dtw.m

[ref,fs] = wavread(‘Menu.wav’);

[test,fs] = wavread(‘T_Menu.wav’);

%ref  = silence(ref);

%test = silence(test);

%whos

[d1,d2x] = dtw(ref,test,fs);

figure(101);

%subplot(211);

plot(ref);

hold on

plot(test,’k');

title(‘Before DTW’);

legend(‘Ref Signal’,'Test Signal’)

figure(102);

%subplot(212);

plot(d1);

hold on

plot(d2x,’k');

%axis([ 0 20000 0 1]);

title(‘After DTW’);

legend(‘Ref Signal’,'Test Signal’)

Figure 2(a): Plot of the reference signal

Figure 2(b): Plot of the test signal

Figure 2(c): Plot of the test signal after DTW with reference signal

This can be achieved by using this equation:

input = input/max(abs(input));

Example:

***********************************************************

%normal.m

function [n] = normal(sound);

n = sound/max(abs(sound));

%A = A/max(abs(A));

***********************************************************

%main_normal.m

[A,fs] = wavread(‘Nadzirah.wav’);

m = normal(A);

figure(1);

subplot(211);

plot(A);

title(‘Before normalization’);

ylabel(‘Amplitude’)

subplot(212);

plot(m);

title(‘After normalization’);

xlabel(‘Time,s’);

ylabel(‘Amplitude’);

Figure 1: plot of Anne.wav before and after normalization

Figure2: Plot of Calendar.wav before and after normalization

Update:

function [n] = norm(sound);

n = sound/max(abs(sound));

A low-pass filter is a filter that passes low-frequency signals but attenuates (reduces the amplitude of) signals with frequencies higher than the cut off frequency. The actual amount of attenuation for each frequency varies from filter to filter. We can say that low-pass filter plays big role in getting rid of noise and silence that usually appear in the passband and stopband.

Low-pass filter can be seen in two category; FIR (Finite Impulse Response and IIR (Infinite Impulse Response). And in FIR itself, it has three designs that can be implemented which are the REMEZ, FIRLS, and KAISER. Design for IIR varies more than FIR which is Butterworth IIR Filter, Chebyshev Type I IIR Filter, Chebyshev Type II IIR Filter, and Elliptic IIR Filter.

While there are many advantages and disadvantages from FIR and IIR, IIR is seldom used in most processing because IIR is recursive where there is feedback involved. A particular sample from a signal will have multiplications and summations performed. IIR will forever affect the outcome of the filter to one degree or another. The impulse response keeps going. IIR also can be designated to accurately simulate analog filter responses.

In contradiction to IIR, FIR is non-recursive as there is no feedback involved. Any signal sample running through an FIR filter will have a finite number of operations performed on it. Impulse response will finally reach zero.

Theoretically, it is best to implement Butterworth filter. This is mainly because the output will have no ripple (monotone) in passband and stopband to be compared with other IIR filter.

Example coding:

[a,Fs] = wavread(‘Anne.wav’);

A = normal(a);

C = silence(A);

%C = C(1:20000,1);

%fourier transform of wave

fftc = fft(C,512);

w = (0:255)/256*(Fs/2);

Rp = 3;

Rs = 60;

Wp = 8000/(Fs/2);

Ws = 9000/(Fs/2);

%to design low pass filter

[Nb,Wb] = buttord(Wp,Ws,Rp,Rs);

[r,s] = butter(Nb,Wb);

[Hb,wb] = freqz(r,s,512);

figure(1);

plot(wb*Fs/(2*pi),abs(Hb));%plot of the filter

%to filter noise

x = filter(r,s,C);

fftb = fft(x,512);

figure(2);

subplot(311);

plot(C);

subplot(312);

plot(w,abs(fftc(1:256)’));

subplot(313);

plot(w,abs(fftb(1:256)));

Figure 1: Low pass filter

Figure 2(a): Plot of the original wave

Figure 2(b): Plot of the Fourier Transform of the wave

Figure 2(c): Plot of the fft of filtered signal

The signal analysis will be achieved in both time domain and frequency domain.
1. We will approach the time domain processing by studying the pitch extraction method and direct matching processing.
The pitch extraction method in time domain may includes several methods which are Parallel processing method, Data reduction method, or Zero crossing method. Time lagged will also be considered in this process. One of the direct matching process is the template matching process. This may be used where the template of the recorded speech will be compared directly with the template of the stored information by a certain matching algorithm.

2) The frequency domain processing will be done by studying various possible solutions. The speech will first be transformed into frequency domain. This will be achieved through the Fast Fourier transform. The formants of the frequency spectrum or the phonemes will be extracted and compared to the stored phonemes by applying dynamic time warping or stochastic modeling (Hidden Markov Model).

This project will be done using MATLAB as a medium to extract, evaluate and deliver a result for speech recognition system.