Browse Prior Art Database

Speech Recognition Poisson-Based Model for Computing Word Votes

IP.com Disclosure Number: IPCOM000036447D
Original Publication Date: 1989-Sep-01
Included in the Prior Art Database: 2005-Jan-29
Document File: 3 page(s) / 28K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

A technique is described whereby performance of speech recognition is improved through the use of a Poisson-based model and used in the computation of word votes during a polling fast-match operation.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 53% of the total text.

Page 1 of 3

Speech Recognition Poisson-Based Model for Computing Word Votes

A technique is described whereby performance of speech recognition is improved through the use of a Poisson-based model and used in the computation of word votes during a polling fast-match operation.

In prior art, speech recognition systems used a slow and expensive process to compare an unknown utterance with a hypothesized word drawn from an allowable vocabulary. As the vocabulary increased in size, it became increasingly important to first draw up a short list of candidate words from the vocabulary. This was done so that every word in the vocabulary would not have to be compared with the unknown word.

One method used in obtaining a short list of words was to use a polling fast- match method. This was typically used with any speech recognition system having a labeling acoustic processor, such as an acoustical processor which converts an utterance into a sequence of labels. Here, each label casts a varying vote for every word in the vocabulary. By summing the votes across all labels in an unknown utterance, a score was obtained for every word in the vocabulary. The words with the highest scores are taken to be the most likely candidates for the true word.

The length and accuracy of the short lists obtained in the polling fast-match method were dependent on the quality of the votes. In a vote of label "L" for word "W" it is the log probability that a randomly selected label from word "W" will be "L". This log probability was computed from the expected frequency distribution of labels for word "W". Because this definition of votes ignores those labels which do not occur in the unknown utterance, it is considered necessary to introduce the idea of penalties, where words are penalized whenever any of their expected labels did not occur. However, the application of penalties and keeping track of which labels occurred and which did not occur seri...