Browse Prior Art Database

Improved Tree Based Fast Match for Speech Recognition

IP.com Disclosure Number: IPCOM000108394D
Original Publication Date: 1992-May-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 1 page(s) / 43K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

This invention relates to improvements in the tree-based acoustic fast match (1) sometimes used in speech recognition systems. The improvements are obtained by incorporating leafemic statistics (2). into the fast match in addition to the phonetic statistics that are normally used.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 82% of the total text.

Improved Tree Based Fast Match for Speech Recognition

       This invention relates to improvements in the tree-based
acoustic fast match (1) sometimes used in speech recognition systems.
The improvements are obtained by incorporating leafemic statistics
(2). into the fast match in addition to the phonetic statistics that
are normally used.

      The main steps of the new procedure are listed below.
1. Phonetic baseforms for each word in  the  vocabulary  are
converted to leafemic baseforms using the method explained in (2).
2. Using the leafemic baseforms, a leafeme tree is constructed in a
manner analogous to the phonetic tree in (1).
3. Training is carried out individually for the phonetic baseforms
and the leafemic baseforms. Note that each leafeme has a fenemic
baseform associated with it.
4. Fast match distributions are obtained for each phone and each
leafeme by the method of (1). This is done by assigning to each
feneme the maximum possible probability from amongst all the
distributions that occur in that phone or leafeme.
5. A composite fast match distribution is formed for each leafeme by
combining the leafeme fast match distribution with the fast match
distribution of the underlying phonetic phone for that leafeme.  One
very effective way to do this is to form the average of the two
distributions.
6. A tree fast match is performed as in (1), but using the leafeme
tree instead of the phonetic tree, and also using the dynamic
thresholding techniques d...