[Oops, replied privately, resending to the list] 10.03.2014 21:02, Jason Newton wrote: > Small amendments to previous: > > I meant to post this link in the above: > http://www.dsprelated.com/dspbooks/sasp/Overlap_Add_OLA_STFT_Processing.html > Oh, OK, now I see where you are coming from: """ We may now introduce spectral modifications by multiplying each spectral frame $ X_m(\omega)$ by some filter frequency response $ H_m(\omega)$ """ This is unfortunately wrong. I am currently writing a response to the original mail that you wanted to add the link to, and will explain it there (albeit in time domain). > Regarding the FFT size, I think you are talking is the is making sure > the filter sampling is at (filter_size + chunk_size + 1) or higher > when you have the FIR filter defined in the time domain. However the > filter was is defined in the frequency domain and only changes > amplitude of frequency (not phase) so I'm not sure if there can be a > problem or not. Most FIR literature is about applying something > defined in the time domain and examples of non IIR equalizers are > pretty much nonexistent so I fully admit to being in the grey here. I have a quote against this. Please see http://www.dspguide.com/ch17/1.htm and search for "Why isn't it possible to directly use the impulse response shown in 17-1b as the filter kernel?". Read three paragraphs including and below that phrase, and you'll find out that the whole phrase "the filter is defined in the frequency domain" is always wrong. And I am not only talking about "making sure the filter sampling is at (filter_size + chunk_size + 1)" (which is my point 2), but also about making filter_size reasonable (which is point 3). -- Alexander E. Patrakov
