Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!pipex!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!ux1.cso.uiuc.edu!usenet.ucs.indiana.edu!ectl-sub@snowhite.cis.uoguelph.ca
From: ectl-sub@snowhite.cis.uoguelph.ca (Electronic Communal Temporal Lobe  -  ECTL)
Subject: ECTL 2.10.2
Message-ID: <9210191920.AA25727@snowhite.cis.uoguelph.ca>
Sender: <ectl-sub@snowhite.cis.uoguelph.ca>
Organization: Indiana University
Date: Mon, 19 Oct 1992 19:20:56 GMT
Lines: 55


		Electronic Communal Temporal Lobe
		=================================



Issue: 2.10.02

========================================================================

(1) Sid Segalowitz, Brock University, Canada; ssegalow@spartan.ac.brocku.ca
       Subject: Response to 2.10.01; re: Performance Statistics Problem

========================================================================

1)

RE aggregating recognition results from nonindependent sets of
speakers
 
     If I understand your setup correctly, you are sampling 200
training voices with replacement from the total of 240 voices, and
then testing the result on the remaining 40 voices.  You cannot
treat the 6 separate samplings as independent since they clearly
are not.  Any two overlap by at least 80% (160/200).  The standard
probability tables are built on the assumption of independence of
sampling and there may be no way to correct for this.  The degrees
of freedom will be much too high.  Perhaps you have to do
confirmatory model testing with 100 training voices and 20
recognition voices (build the model with the first half and test it
with the second half).  
     I have tried to find the answer to a similar statistical
problem: Correcting Pearson correlation signficance tables in a
situation where the X2 is not independent of the X1 nor X3, and so
on.  My attempt to find a "correction" for the degrees of freedom
was in vain and several statisticians were not able to help.  If
you know of any solution, please let me know.

=========================
End.


+-----------------------------------------------------------------------+
|									|
|		     Electronic Communal Temporal Lobe			|
|									|
|  Moderator:  David Leip, University of Guelph, Computing & Info. Sci. |
|		Guelph, Ontario, Canada  N1G 2W1.   			|
|		(519) 824-4120  ext.3709 (office) ext.4297 (lab).	|
|									|
|  Send submissions for ECTL to: ectl-sub@snowhite.cis.uoguelph.ca	|
|  Send all other requests to: ectl-request@snowhite.cis.uoguelph.ca	|
|  Anonymous ftp archive at: snowhite.cis.uoguelph.ca  in: pub/ectl     |
|									|
+-----------------------------------------------------------------------+
