Subject: Re: About importance of "phase" in sound recognition From: Laszlo Toth <tothl@xxxxxxxx> Date: Wed, 6 Oct 2010 10:20:00 +0200 List-Archive:<http://lists.mcgill.ca/scripts/wa.exe?LIST=AUDITORY>On Tue, 5 Oct 2010, emad burke wrote: > I'm exactly talking about what is called "in-sensitivity to phase". I'm > talking about the phase information that is > discarded in the process of MFCC feature extraction and it has been proven > to be succesfull feature set for speech recognition. As I said, the problem is that quite probably different things are meant by "phase" in the classic "insensitivity to phase" claim and the speech technology interpretation. In speech recognition FFT is applied to decompose the signal into frequency components. FFT decomposes the signal into sine waves with GIVEN frequencies. For example, if the signal consists of only one sine wave, but its frequency is not among the FFT frequencies, then you sine wave gets decomposed into a lot of sines, and the phase information will guarantee that after summing these components, you receive back just one sine wave. If you change the phases, the whole thing falls apart. I don't exactly know what the old sentence "the ear is insesitive to phase" meant by phase, but surely not the phases of the components obtained by FFT. So I think this if the main source of confusion. > On the other hand, couple of years ago there was a > publication by a mathematician (pete-cassaza) that kind of reinforced the > argument of phase insensitivety of speech recognition, but this time > mathematically; very briefly stating that if you have a redundant set of > magnitude coeeficients, then phase doesnt matter at all, This is stated too briefly. What is the mathematical meaning of "doesn't matter"? Do you have a reference? Laszlo Toth Hungarian Academy of Sciences * Research Group on Artificial Intelligence * "Failure only begins e-mail: tothl@xxxxxxxx * when you stop trying" http://www.inf.u-szeged.hu/~tothl *