Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!news.alpha.net!uwm.edu!uwvax!sinetnews!news.u-tokyo.ac.jp!news.tisn.ad.jp!kddlab!peter
From: peter@multix3.elb.lab.kdd.co.jp (Peter SCHNEIDER)
Subject: HMM decomposition
Nntp-Posting-Host: 133.128.14.77
Sender: news@kddlab.lab.kdd.co.jp (USENET News System)
Organization: KDD R&D Laboratories, Kamifukuoka, Japan
Date: Wed, 1 Mar 1995 02:44:41 GMT
Message-ID: <PETER.95Mar1114441@multix3.elb.lab.kdd.co.jp>
Lines: 26


I am doing research on improving the recognition performance of a
continuous speech recognizer in the presence of colored, narrowband,
non-stationary noise.

To deal with that problem, I found HMM decomposition very promising, but of
course the big question is how to compute the output probability of a state
in the combined model (I am using MFCC).

The approximation methods I found proposed in some papers are not
appropriate for my case, and the most recent papers I saw were from 1992,
has this approach been given up?

I would very much appreciate some pointers to recent work in HMM
decomposition or, even better, if someone could tell me about his
experience in that field. Also, if I am on an utterly wrong track and there
is actually a much better solution to my noise problem, please let me know.

Thanks for your help.
 
--
_____________________________________________________________________
  Peter SCHNEIDER            KDD R&D Laboratories, Multimedia Group
  email  peter@multix3.elb.lab.kdd.co.jp        tel +81 492 66 7390
  www    http://www.cica.fr/~schneide           fax +81 492 66 7510
_____________________________________________________________________
