Browse Prior Art Database

Thresholding Scheme for Fast Labeling

IP.com Disclosure Number: IPCOM000100998D
Original Publication Date: 1990-Jun-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 2 page(s) / 65K

Publishing Venue

IBM

Related People

Bahl, L: AUTHOR [+6]

Abstract

This invention is a modification of the algorithm that was introduced in (1). It provides a significant improvement both in the reduction of the labeling time of discrete parameter Markov model speech recognition systems and in the decoding accuracy. These advantages were achieved by using a thresholding scheme in the process of constructing a confusion matrix for labels from the training data.

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

Thresholding Scheme for Fast Labeling

       This invention is a modification of the algorithm that
was introduced in (1).  It provides a significant improvement both in
the reduction of the labeling time of discrete parameter Markov model
speech recognition systems and in the decoding accuracy.  These
advantages were achieved by using a thresholding scheme in the
process of constructing a confusion matrix for labels from the
training data.

      A description of the labeling method that attaches labels to
spectral vectors and the motivation for a faster labeling procedure
are given in (1).  Other situations, where the speed of the decoding
procedure (that includes also the labeling procedure) is of
exceptional importance, are given in (2).

      The quantization of the acoustic data that is produced by a
speaker, i.e., representation of this data in a discrete form (as
strings of labels) results in a loss of information and subsequently
leads to decoding errors.  The labeling method that was suggested in
(1) provided a reduction in the labeling time but not an improvement
in the decoding accuracy.  This happens because in the method in (1)
there is no distinction between information about typical and rare
occurrences of labels that are produced from the training data.  In
this invention, we provide a modification of the algorithm from (1)
that makes a distinction between various entries of the confusion
matrix depending on how typical they are for a given speaker.  This
leads us to a significan...