CRG-TR-89-5 available

[Carol Plathan]

       PLEASE DO NOT FORWARD TO OTHER NEWSGROUPS OR MAILING LISTS
       **********************************************************

The following paper was presented at a recent meeting of the Acoustical
Society of America:

		Context-Modulated Discrimination of Similar Vowels
		     Using Second-Order Connectionist Networks

				R. Watrous

		Dept. of Computer Science, University of Toronto
			 and Siemens Corporate Research


Discrimination of two vowels in the context of voiced and unvoiced stop
consonants is investigated using connectionist networks. Separate
discrimination networks were generated from samples of the vowel centers of
[e,ae] for the six contexts [b,d,g,p,t,k] for one speaker. A single
context-independent network was similarly generated.  The context-specific
error rate was 1%, whereas the context-independent error rate was 9%.  A
method for merging isomorphic context-specific networks into a single network
is described, that uses singular value decomposition to find a minimal basis
for the set of context-specific weight vectors. Context-dependent linear
combinations of the basis vectors may then computed using second-order network
units. Compact networks can thus be obtained in which the vowel discrimination
surfaces are modulated by the phonetic context. In one experiment, as the
number of basis vectors was reduced from 6 to 3, the error rate increased from
1% to 3%.  A context-modulated network with three basis vectors and a
context-independent network were also trained on the same data using standard
methods of nonlinear optimization.  The discrimination error rate using the
context-independent network was as low as 2.6%, whereas the context-specific
recognition error rate was as low as 0.3%. It is concluded that compact
context-sensitive connectionist networks which result in very high phoneme
discrimination accuracy can be constructed and trained.

---------
To obtain a copy of this paper please send your real mailing address to 
carol at ai.toronto.edu
---------