Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!warwick!pipex!uunet!haven.umd.edu!darwin.sura.net!howland.reston.ans.net!xlink.net!fauern!lrz-muenchen.de!regent!media!alfred
From: alfred@lis.e-technik.tu-muenchen.de (Alfred Hauenstein, Mitarbeiter LIS)
Subject: Re: VQ Codebook for Speech Recognition
Message-ID: <CA5I4y.2no@regent.e-technik.tu-muenchen.de>
Sender: news@regent.e-technik.tu-muenchen.de (News System.)
Reply-To: A_Hauenstein@lis.e-technik.tu-muenchen.de
Organization: TU Muenchen / LIS
References: <1993Jul12.075431.8414@ericsson.se>
Date: Wed, 14 Jul 1993 10:50:58 GMT
Lines: 28

In article 8414@ericsson.se, ebcguss@ebcw196.ericsson.se (BO/EBC/B/XE Claes Gussing  08-682 4978) writes:
> For a speaker dependent pattern recognition system I will need a codebook for my
> vector quantization. Should I build this codebook for each user, since the system
> will be speaker dependent, or should I use the same codebook for all of them; in the
> latter case, is there any good codebooks available someplace - i.e. one for male and
> one for female voice, english language?

If you want to build a speaker dependent system, there is only one advise:
Use different codebooks for each speaker, because this reduces the error rat
for recognition. (As a rule of thumb: by a factor of 2).

	Alfred


--------------------------------------------------
!!! Please notice the new 5-digit zip-code !!!
!!! (valid since July, 1st, 1993)          !!!

Alfred Hauenstein
Chair for Integrated Circuits (Prof. Ruge)
Technical University of Munich
Arcisstr. 21
80290 Muenchen
GERMANY
Phone:	++49 89 231817-47
Fax:	++49 89 231817-30
e-mail:	A_Hauenstein@lis.e-technik.tu-muenchen.de

