wav2ftr1.m

最全的ASR工具集合

function parameter= wav2ftr1(Inwav)
% WAV2FTR Wave to MFCC feature extraction

setparam;
if isstr(Inwav),
	[Inwav,Fs,NBits] = wavread(Inwav);
end

% ====== Step 1: Pre-emphasis.
Inwav = filter([1, -0.95], 1, Inwav);

% ====== Step 2: frame blocking.
Frame = buffer(Inwav, frameSize, overlap);
load tri_coef.mat   % for triBandFilter parameter : fstart,fcent,fstop.

normalize_coff = 10;
energy = sum(Frame.^2)/frameSize;
index = find(energy < threshold);
energy(index) = [];
logEnergy = 10*log10(energy)/normalize_coff;
Frame(:, index) = [];
parameter = [];
for i = 1:size(Frame, 2);
      % ====== Step 3: hamming window.
      %            w(m) = 0.54 - 0.46 cos(2*pi*m / n) , 1<=m<=n+1.
      %Wframe = Frame*(0.54 - 0.46*cos(2*pi*i / (FrameCount-1)));
      Wframe  = hamming(frameSize).*Frame(:,i);
      
      % ====== Step 4: fast fourier transform.
      %            Using FFT function to calculate. 
      %            Compute square of real part and imaginary part.
      %fftFrame = sqrt(real(fft(Wframe)).^2+imag(fft(Wframe)).^2);
      fftFrame = abs(fft(Wframe));
      
      % ====== Step 5: triangular bandpass filter.
      %            Using user defined function triBandFilter(fftFrame{i}).
      P = 20;      %P means counts of log spectral magnitude.
      tbfCoef = triBandFilter(fftFrame,P,fstart,fcent,fstop);
      %tbfCoef = bandfilt(fftFrame,P,fstart,fcent,fstop);
      
      % ====== Step 6: cosine transform.
      %            Using DCT to get L order mel-scale-cepstrum parameters.
      L = 12;      %L means L order , generally L is 12.
      cepstrum = mel_cepstrum2(L,P,tbfCoef);
      parameter = [parameter cepstrum'];
end;
parameter = [parameter; logEnergy];

%=========compute delta energy and delta cepstrum============
%Calculate delta cepstrum and delta log energy
%Combine them with cepstrum and log energy, get 26 order parameter.
delta_window = 2;
parameter = getDeltaFeature(delta_window, parameter);

%========================Subfunction================================

% ====== Triangular Band Filter
function tbfCoef = triBandFilter(fftFrame,P,fstart,fcent,fstop)
%The function is triangular bandpass filter.
for i = 1 : P,
   for j = fstart(i) : fcent(i),
      filtmag(j) = (j-fstart(i))/(fcent(i)-fstart(i));
   end;
   for j = fcent(i)+1: fstop(i),
      filtmag(j) = 1-(j-fcent(i))/(fstop(i)-fcent(i));
   end;
   tbfCoef(i) = sum(fftFrame(fstart(i):fstop(i)).*filtmag(fstart(i):fstop(i))');
end;
tbfCoef = log(tbfCoef.^2);

% ====== Mel-scale cepstrum
function cepstrum = mel_cepstrum2(L,P,tbfCoef)
%compute mel-scale cepstrum , L should be 12 at most part.
for i=1:L,
	coef = cos((pi/P)*i*(linspace(1,P,P)-0.5))';
	cepstrum(i) = sum(coef.*tbfCoef');
end;

% ====== Delta cepstrum and delta log energy
function parameter = getDeltaFeature(delta_window,parameter)
% Compute delta cepstrum and delta log energy.
rows  = size(parameter,1);
cols  = size(parameter,2);
temp  = [zeros(rows,delta_window) parameter zeros(rows,delta_window)];
temp2 = zeros(rows,cols);
denominator = sum([1:delta_window].^2)*2;
for i = 1+delta_window : cols+delta_window,
   subtrahend = 0;
   minuend    = 0;
   for j = 1 : delta_window,
      subtrahend = subtrahend + temp(:,i+j)*j;
      minuend    = minuend + temp(:,i-j)*(-j);
   end;
   temp2(:,i-delta_window) = (subtrahend - minuend)/denominator;
end;
parameter = [parameter ; temp2];