Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!fs7.ece.cmu.edu!hudson.lm.com!news.pop.psu.edu!news.cac.psu.edu!howland.reston.ans.net!vixen.cso.uiuc.edu!uwm.edu!caen!msunews!harbinger.cc.monash.edu.au!yoyo.aarnet.edu.au!frodo.cc.flinders.edu.au!
From: egssr@flinders.edu.au
Subject: How much data needed for Discrete HMM?
Message-ID: <1995Apr2.081849.24193@frodo.cc.flinders.edu.au>
Sender: @frodo.cc.flinders.edu.au
Date: Sun, 2 Apr 1995 08:18:49 GMT
Organization: Flinders University of South Australia
X-Newsreader: <WinQVT/Net v3.9>
Keywords: HMM
Lines: 30

Dear Speech Experts,

  I am trying to develop a small vocabulary isolated word speaker
  dependent speech recognition system.

  I am using discrete HMMs to model each word in the vocabulary.
 
I was wondering how many examples of each word are needed to achieve a 
  "robust" representation of each word.  I am using the Baum-Welch 
  algorithm with scaling to train the HMMs but I will be comparing the
  results with HMMs trained with the simpler Viterbi optimal path 
  algorithm.

  Also, how many examples are needed to generate a good codebook
  for the vector quantiser.

  The feature vector I am using comprises 12 LPC derived cepstral
  coefficients plus 12 delta cepstral coefficients.

  I would appreciate it if someone could email me these answers to
  egssr@flinders.edu.au


Thanks in advance

Suneel Randhawa

PS. If any of these questions are already detailed in the FAQ, I would
    appreciate it if someone could direct me to it. In this case I 
    apologise for the inconvenience. 
