It seems you forgot to do the 50% overlap. Also, if you intend on actually modifying the signal between the fft and the ifft, it's best to apply the window both before the fft and after the ifft. You need a "power complementary" window for that. The simplest is sqrt(Hanning), which is just a sine window. Note that you *can't* use a Hamming window in any case.

It is willfully that I "forget" overlap for testing the effect of windowing. The difference between the 2 turns is windowing only. But at the 1st turn, the reconstructed signal sound well,in contrast with the 2nd turn bad effect. It must be windowing which corrupt the signal.

Quote from: hyeewang on 2008-12-04 02:03:59

It is willfully that I "forget" overlap for testing the effect of windowing. The difference between the 2 turns is windowing only. But at the 1st turn, the reconstructed signal sound well,in contrast with the 2nd turn bad effect. It must be windowing which corrupt the signal.

You're multiplying your signal times a periodically extended hanning function.  This should change the way the signal sounds quite a lot.  Were you expecting something else?

What is it that you want to do?

Since ifft(fft(x))=x you don't need to think about the FFT to answer the first question.

Regarding window functions: Yes, if you intend to modify the "FFT data" somehow you should use a window prior FFT and after iFFT with 50% overlap and an appropriate window like jmvalin suggested. But neither hann, hanning, hamming, nor the square-root versions of these are appropriate.

Try Vorbis' window:

function w = vorbiswin(n)
w = sin((pi/2).*sin(pi.*((0.5:n)'./n)).^2);
return;

If you just want to filter your signal (as in convolution) you might want to drop the whole windowing thing. There's no need for it in case you know exactly the filter's impulse response's length.

Cheers!
SG

I would ask the same question already asked: What are you talking about?

When you are talking about percentage of overlap I guess you are calculating Short Time Fourier Transform (STFT) or Time-Dependant Fourier Transform. In this case using window functions with Constant Over-Lap-Add Property and accordingly a proper overlap you can calculate a sampled approximation to the Discrete Time Fourier Transform of the analyzing signal and of course to analyze the signal. When using Inverse Fourier Transform you can get as many time points as frequency bins are there in the spectrum. So, not an infinite long signal.

When you are talking about a “reconstructing of speech signal” I guess you need to reconstruct a long signal.
In a case you want to process an infinite long signal with a finite filter (SebastianG mentioned this) then, I think, another techniques are used: overlap-add and overlap-save methods. This approach is described very well in Discrete-Time Signal Processing by Alan V. Oppenheim and Ronald W. Schafer and the implementation is straightforward.
If you have questions about that most probably I could answer them.

When you modify the transform, it is the same as convolving the signal with the resulting time-domain signal from the ifft of the modification signal.

You need enough overlap, which includes zeros before and after the analysis window, so that you can overlap by enough to account for the actual time-domain length of the effective impulse response you are creating in the frequency domain.

You can use a Hann window on the analysis side, and no window on the reconstruction side, just overlap-add, if your second signal is appropriately conditioned, too.

But you did not tell the principle of why using overlap-add.

My apologies if it's somewhat off-topic, but what are the benefits and drawbacks to using a DCT instead of FFT in this sort of application?