Browse Prior Art Database

Replacing Singular Distributions with Locally Optimal Distribution in Zuelogical Training for Tangora

IP.com Disclosure Number: IPCOM000106771D
Original Publication Date: 1993-Dec-01
Included in the Prior Art Database: 2005-Mar-21
Document File: 2 page(s) / 55K

Publishing Venue

IBM

Related People

Bellegarda, J: AUTHOR [+2]

Abstract

In order to perform k-means training in an n-dimensional metric space using diagone requires non-singular seeds. In Zuelogical training for the Tangora Automatic Speech Recognition System, singular seeds are sometimes created. Disclosed is a way to replace the singular seed with a locally-optimal non-singular seed.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Replacing Singular Distributions with Locally Optimal Distribution in Zuelogical Training for Tangora

      In order to perform k-means training in an n-dimensional metric
space using diagone requires non-singular seeds.  In Zuelogical
training for the Tangora Automatic Speech Recognition System,
singular seeds are sometimes created.  Disclosed is a way to replace
the singular seed with a locally-optimal non-singular seed.

      In order to perform k-means training on n-dimensional vectors
using diagonal Gaussian prototypes, one requires non-singular seeds.
In Zuelogical training for the Tangora system, singular seeds are
sometimes created.  There are several options one could employ to
solve this problem.  One could throw out the seed and replace it with
another.  Another option is to make the singular seed non-singular by
binning more vectors until the seed becomes non-singular.  But these
two options are to be used carefully, as they could lead to the
dilution of important, but rare, acoustic events.  Yet a third
approach is to keep the mean component of the seed unmodified, but
replace the singular variances with non-singular variances gotten by
binning more vectors until the seed becomes non-singular.  This is
more reasonable since the mean component can usually be estimated
with adequate precision, even for a small number of vectors.

      The third option mentioned above, replacing the singular
variance component but preserving the mean componen...