Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!pipex!uunet!elroy.jpl.nasa.gov!ames!han.hana.nm.kr!ccsun2.sogang.ac.kr!news
From: vibrato@ailab1.sogang.ac.kr (Yang JoonYong)
Subject: [Q] Viterbi in HMM
Message-ID: <1993Oct12.073005.2457@ccsun2.sogang.ac.kr>
Sender: news@ccsun2.sogang.ac.kr
Organization: Sogang University
X-Newsreader: TIN [version 1.1 PL9]
Date: Tue, 12 Oct 93 07:30:05 GMT
Lines: 22


Hi, all

I have been experimenting with VQ based HMM system similar to that described in the 1983 Bell paper(Rabiner et al). 
 
At the beginnging, I let the self-transition of the last state == 1.0 as paper, and I noticed is that state sequences found with the Viterbi algorithm seem to spend very much time in the last state of the model.  

So, I have let the transition probability to be 0, and have got recognition rate a little lower than that before. 

I'm wandering that last self transition probability is to 0.   
According to Rabiner's paper, 
      for each i, sigma(Aij) must be 1.    
                   j=1..N
If it is wrong what I did, how to define a transtion of last state in order to avoid spending much time at the last state ? 
Thanks a lot for your help..
   

                            Yang Joon Yong.
     

E-mail: vibrato@ailab1.sogang.ac.kr  

