Newsgroups: comp.speech
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!zombie.ncsc.mil!news.duke.edu!news-feed-1.peachnet.edu!gatech!rutgers!news.cs.indiana.edu!fcummins@cs.indiana.edu
From: "Fred Cummins" <fcummins@cs.indiana.edu>
Subject: 5th order PLP: how many cepstral coeffs?
Message-ID: <1995Mar22.192404.26016@news.cs.indiana.edu>
Sender: root@news.cs.indiana.edu (Operator)
Organization: Computer Science, Indiana University
Date: Wed, 22 Mar 1995 14:23:58 -0500
Lines: 25

Question to those versed in cepstral coeffecients and/or Hynek
Hermansky's PLP representation:  

A 5th order PLP is found to be more or less optimal for speech.  How
many predictive coefficients should this produce?  If they are
cepstral coefficients (as the public domain code seems to generate),
is the number of output features (i.e. cepstral coefficients)
logically independent of the order of the model?  The code will
produce output for such apparent nonsense as 5th order model, 15
output features (though automatic speech recognition is not improved
by including these additional "features").

Thanks

Fred

-- 
Fred Cummins		| fcummins@cs.indiana.edu 
Lindley Hall 406,	| (812) 855-8702 (office) 
Indiana University,	| 
Bloomington, IN 47405   | http://www.cs.indiana.edu/hyplan/fcummins.html
____________________________________________________________
Never give a monster the work of an eee-vill scientist......
------------------------------------------------------------         

