Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!news.alpha.net!uwm.edu!news.doit.wisc.edu!decwrl!tribune.usask.ca!quartz.ucs.ualberta.ca!news.ucalgary.ca!hill
From: hill@salab1.psych.ucalgary.ca (David Hill)
Subject: Re: How to segment speech signals
Message-ID: <Dec21.061622.64111@acs.ucalgary.ca>
Date: Wed, 21 Dec 1994 06:16:22 GMT
References: <3d7g0k$det@math.mps.ohio-state.edu>
Organization: Paranoids United
Lines: 47

In article <3d7g0k$det@math.mps.ohio-state.edu>,
lmr <ren@math.ohio-state.edu> wrote:
>As a new comes in DSP, I need help about how to segment speech signals
>(in wave form). What is the most effective way to tell where a vowel
>starts and stops? the most popular one?
>
>
>Any pointers about code, literature or anything are greatly
>appreciated!!
>
>
>-- 
>Liming Ren                       |
>Dept. of Mathematics             |  231 West 18th Avenue
>The Ohio State University        |  Columbus, Ohio43210

As Ken Stevens said many years ago, the "segmentation problem" in speech is
the problem that you can't (segment, that is).  Speech is mostly continuous,
and the approved mthods of segmenting all mostly involve slicing up the speech
into short segments (the salami approach), and then deciding which of these
may most reasonably be concatenated (or ignored) to form something akin
to phones (not "phonemes" -- those are the names given to categories of
sounds that may not even be similar).  This is basically what
Hidden Markov Modelling (HMM) and Dynamic Programming (DP) approaches do,
but no-one has really soleved the continuous speech recognition problem
yet.

Personally, I prefer to avoid segmenting altogether, and look for phonological
events, and use those as evidence on which to base unit recognition
decisions.  The question then is, what units do you think should be recognised.
There's no law which says that you have to recognise phones (or phoneme
categories).  Its just that people think it will be easier to handle
40 or so symbols, than try to deal with syllables or words or ...

Still, there are Isolated Word Recognisers (IWRs) and they often recognise
phrases too, but that's not really what I meant.

Welcome to the Grail Quest!

david :-)
------------

-- 
David R. Hill, CS Dept., U. Calgary         | Imagination is more
Calgary, AB, Canada T2N 1N4 Ph: 403-220-6315| important than knowledge.
hill@cpsc.ucalgary.ca       Fx: 403-282-6778|         (Albert Einstein)
NeXTMail: hill@trillium.ab.ca (Preferred)   | Kill your television!
