Browse Prior Art Database

Procedure for Using Contextual Information to Obtain Improved Estimates of Word Probabilities in a Speech Recognition System

IP.com Disclosure Number: IPCOM000038491D
Original Publication Date: 1987-Jan-01
Included in the Prior Art Database: 2005-Jan-31
Document File: 2 page(s) / 61K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+5]

Abstract

The present invention discloses methodology for incorporating additional context into m-gram language models used in speech recognition without exponentially increasing the number of m-grams which results from increasing m. In a tri-gram language model function words -- such as "the", "a", "of" -- are well predicted from the previous two words. Content words -- such as "computer", "wealthy", "hire" -- often require more context to be predicted accurately. Because the number of m-grams increases exponentially with m, it is not practical to simply increase m to obtain more context. To achieve the increased context without an attendant combinatorial explosion, the present invention produces an estimator for predicting the last word in an m-gram (or, as described herein, a tri- gram).

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 52% of the total text.

Page 1 of 2

Procedure for Using Contextual Information to Obtain Improved Estimates of Word Probabilities in a Speech Recognition System

The present invention discloses methodology for incorporating additional context into m-gram language models used in speech recognition without exponentially increasing the number of m-grams which results from increasing m. In a tri-gram language model function words -- such as "the", "a", "of" -- are well predicted from the previous two words. Content words -- such as "computer", "wealthy", "hire" -- often require more context to be predicted accurately. Because the number of m-grams increases exponentially with m, it is not practical to simply increase m to obtain more context. To achieve the increased context without an attendant combinatorial explosion, the present invention produces an estimator for predicting the last word in an m-gram (or, as described herein, a tri- gram). Specifically, the invention teaches two alternative approximations which use bi-grams and tri-grams together to "increase" context. It is assumed that, based on accumulated data, tri-gram probabili- ties Pr(Wo W1,W2) have been pre-computed. It is also assumed that bi-gram probabilities Pr(W Wo) have been pre-computed. Pr(W Wo) denotes the probability that a randomly selected word from a window extending 3TN words prior to Wo will be word W, (N >
2). The above word-trigram and window-bigram probabilities may be obtained from a large body of training text using the methods described in [*]. Let Wi denote the w...