Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!pipex!uunet!seas.gwu.edu!marshall
From: marshall@seas.gwu.edu (Christopher Marshall)
Subject: Re: not phonemes, but.. _______?
Message-ID: <1993Jan29.212833.7575@seas.gwu.edu>
Sender: news@seas.gwu.edu
Organization: George Washington University
References: <RG.93Jan26210205@nymph.msel.unh.edu> <gel.728332449@zen.sys.uea.ac.uk>
Distribution: comp
Date: Fri, 29 Jan 1993 21:28:33 GMT
Lines: 36

In article <gel.728332449@zen.sys.uea.ac.uk> gel@sys.uea.ac.uk (Gareth Lee) writes:
>rg@msel.unh.edu (Roger Gonzalez) writes:
>
>>What is the term for the unit of speech that takes into account the preceeding
>>and trailing sounds?  I seem to recall reading something about that using this
>>method instead of phonemes produced much more normal sounding speech, and that
>>there were still a reasonable (300-400?) number of these units in the english
>>language.
>
>Triphones probably??  These are just concatonated phoneme triplets, ie,
>
>P1|P2|P3 where P2 is in the context of P1 and P3.
>
>The thing that suprises me in you posting is that you say there are 300-400
>used in english.  There will be abot 70^3 possible triphones (350000).
>Whilst a large proportion of these will be illegitimate there are presumably
>several tens of thousand valid triphones.  This is why few people use them
>as a recognition model - they cannot find training sets large enough!!
>
>----------------------------------------+-----------------------------------
>  Dr Gareth Lee                         |  Phone:  +44 603 592878 (direct)
>  Advanced Pattern Processing Group     |  Fax:    +44 603 507720
>  School of Information Systems         |  email:  gel@sys.uea.ac.uk
>  University of East Anglia             |  JANET:  gel@uk.ac.uea.sys
>  Norwich, NR4 7TJ.                     |  finger: gel@mutley.sys.uea.ac.uk
>  United Kingdom.                       |
>----------------------------------------+-----------------------------------

Actually I think there are more like 2000 - 3000 triphones that are
used in a typical large vocabulary (1000 words) task.  The Sphinx system
that Kai Fu Lee worked on uses triphones.  They used deleted interpolation
to compensate for the lack of a larger training set than the TIMIT.

Chris Marshall
marshall@seas.gwu.edu

