About v_overlapadd

[QiangJI123]

Hello, when applying this function, we found that the recovery result has some improvement compared to the clean speech.(i.e. v_enframe ->fft ->ifft -> v_overlapadd ,hamming,256 point,64 inc). And there is a line in the program ：
(z(repmat(1:nf,nr,1)+repmat((0:nr-1)'*inc+rem((0:nr-1)',nb)*no,1,nf))=f.*repmat(w,nr,1);)
why we should f.*repmat(w,nr,1) ?
I look forward to your reply. Thank you.

[voicemike]

Hi,

(1) The expression f.*repmat(w,nr,1) multiplies each frame (i.e. each row of f) by the window w.
(2) The sequence v_enframe ->fft ->ifft -> v_overlapadd should have no effect on the samples except for the (n-inc) samples at both ends. However it is important to use the 'periodic' option of hamming() and also to normalize the window (see the code below). I hope this helps. Mike

inc=64;
nf=4inc;
s=randn(4nf,1); % create a random signal
w=hamming(nf,'periodic'); % Hamming window with period nf
w=w/sqrt(sum(w(1:inc:nf).^2)); % normalize window
f=fft(v_enframe(s,w,inc),nf,2); % do STFT: one row per time frame
x=v_overlapadd(ifft(f,nf,2),w,inc); % reconstitute the time waveform
e=x-s; % calculate reconstruction error
max(abs(e(1+nf-inc:end-nf+inc))) % error is zero except at ends

[voicemike]

Autoformatting seems to have deleted the multiplications in lines 2 and 3 above.