Re: Speech(or Phase) Reconstruction from Magnitude Spectrum

• To: AUDITORY@xxxxxxxxxxxxxxx
• Subject: Re: Speech(or Phase) Reconstruction from Magnitude Spectrum
• From: ismail uysal <uysal_i@xxxxxxxxx>
• Date: Mon, 31 Jan 2005 13:45:02 -0800
• Delivery-date: Mon Jan 31 16:56:34 2005
• Domainkey-signature: a=rsa-sha1; q=dns; c=nofws; s=s1024; d=yahoo.com; b=6o8eRnBsPDgRM6qPJVTE6shm9cCH/cEIxc3374pxYcPeU4R/fsv35ogGuzOANp8U3WI34uaEXGUE6GSUyFUYowsLuRUAvccQpJPurACoXulbNSszFN1SOy+OV8aLSpj8h1XPU1djAHpd9pp/VMzrRe5taIoeLw3/qXKTqxPQvFE= ;
• Sender: AUDITORY Research in Auditory Perception <AUDITORY@xxxxxxxxxxxxxxx>

Hello,

I just read the paper "Signal Estimation from Modified STFT" and before I start to implement this and see the results, I need to ask something to people who had experiences with this subject or paper before.

What I understood is that we assume we have our modified STFT magnitude of the original signal. Then we pick an initial guess, like white gaussian noise for our estimate, and we use an iterative process to minimize the error between the modified STFT and STFT of the estimated signal. My problem is with the iteration part. For our, say second iteration, we keep the phase information from our first guess, make its STFT magnitude equal to modified STFT magnitude, and after we take inverse transform we divide it by the window function. The thing I do not understand is that we don't introduce any change in phase information at any iteration step. So this is like an STFT magnitude estimation using modified STFT magnitude. However, what I would like to have is STFT instead of STFT magnitude.

Is the whole trick in the division by window function, (a modified version of Hamming Window) , or is there something that I'm missing?

Thank you very much again for your help.

Regards,
Ismail Uysal
U of Florida, CNEL

Do you Yahoo!?
Yahoo! Search presents - Jib Jab's 'Second Term'