Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!math.ohio-state.edu!newsfeed.acns.nwu.edu!ftpbox!mothost!schbbs!salaam!assaleh
From: assaleh@mot.com (Khaled Assaleh)
Subject: Re: Energy Measures in Feature Vectors?
Reply-To: assaleh@mot.com
Organization: Motorola, Inc., GSTG
Date: Fri, 5 May 1995 19:05:06 GMT
Message-ID: <1995May5.190506.20487@schbbs.mot.com>
References: <3o8jva$edu@prakinf2.PrakInf.TU-Ilmenau.DE>
Sender: news@schbbs.mot.com (SCHBBS News Account)
Nntp-Posting-Host: 137.144.1.19
Lines: 30

>    I am currently experimenting with isolated word recognition using
>  discrete HMMs.  The main features that I am using are mel scale 
> cepstral coefficients.  Just using these features gives a recognition
> accuracy around 90 percent for a vocabulary consisting of the 26 
> letters of the English alphabet.
> 
> After consulting various texts (Rabiner et al, Deller et al  etc)
> it seems that short term Energy can be used to supplement the mel
> scale coefficients.  However, when I add energy measures (log energy)
> to the feature vector, the recognition accuracy falls by about 5 perecent.
> I have normalised the energy measures according to the methods prescribed
> in the text books yet this doesn't improve recognition accuracy, in
> fact it is less than without using log energy.
> 
> Is it worthwhile pursuing short term energy measures for isolated word
> recognition?

You might want to try Delta-energy not energy!

-Khaled
---
 .................................................................. 
 | Khaled Assaleh, Ph.D.    ||   Secure Communications Section    | 
 | Motorola, GSTG           ||   Speech & Signal Processing Lab.  |           
 | Mail Drop H1175          ||   (SSPL)                           |
 | 8201 E. McDowell Rd.     ||   Voice: (602) 441-1467            |
 | P. O. Box 1417           ||   Fax  : (602) 441-5810            |
 | Scottsdale, AZ 85252     ||   email: assaleh@ziggy.geg.mot.com |
 ``````````````````````````````````'''''''''''''''''''''''''''''''' 

