Newsgroups: comp.speech
Path: lyra.csx.cam.ac.uk!doc.ic.ac.uk!agate!library.ucla.edu!psgrain!nntp.cs.ubc.ca!fornax!bhaskar
From: bhaskar@cs.sfu.ca (Bhaskar Bhattacharya)
Subject: Re: Determining Pitch Period
Message-ID: <1994Aug22.051300.8034@cs.sfu.ca>
Organization: Faculty of Applied Science, Simon Fraser University
References: <338i1pINNltq@life.ai.mit.edu>
Date: Mon, 22 Aug 1994 05:13:00 GMT
Lines: 46

In article <338i1pINNltq@life.ai.mit.edu>,
Robyn Kozierok <robyn@cocoa-chex.ai.mit.edu> wrote:
>
>It is my understanding that in order to do spectral analysis using
>DFTs, you need to first determine the pitch period of the segment
>and then do a DFT on a segment of that length.
>
>(a) is that correct?
>

Depends on what you mean by 'spectral analysis'. In general you can take
any segment of speech, choose an appropriate window and use DFT/FFT to
obtain an estimate of the spectrum. What you said is correct if you want
a harmonic representation of speech with pitch frequency as the fundamental
frequency. 

>(b) is there an easy to determine the pitch period of the segment
>

Not that I know of. All existing algorithms produce wrong pitch values
some time or the other.

>(c) how accurate do you have to be in the determination to get decent
>results from the DFT, and the related question, if the pitch period
>is "close" to a power of two, can you just use the close power of 2
>to do an FFT instead of a DFT?
>

A pitch value in terms of integer number of samples is good enough for most
purposes with speech digitized at 8 KHz.

The pitch does not need to be a power of 2, you can always pad zeros to make it
a power of 2 and then take FFT. Of course you need to pick up the correct 
FFT coefficients to pick up the harmonic magnitudes. Also, the results will
be more sensitive to errors in pitch measurement.

>(d) anything else?
>
>Any help, including pointers to the appropriate literature or textbooks
>would be appreciated.  (Don't point me to the FAQ, I've already read it.)
>

Try to look into the INMARSAT documentation on the IMBE standard for other
ways of doing it.

Bhaskar
