Speech recognition incorporating a priori probability weighting

Details Speech recognition incorporating a priori probability weighting Speech recognition incorporating a priori probability weighting

1997-07-16

1999-12-07

Wieland, Susan

Data processing: speech signal processing, linguistics, language

Speech signal processing

Recognition

704239, G10L 506

Patent

active

059999020

DESCRIPTION:

BRIEF SUMMARY
BACKGROUND OF THE INVENTION

1. Field of the Invention
This invention relates generally to apparatus and method for use in speech recognition.
2. Related Art
The speech recogniser to be described finds application in situations where a recognition process is to be undertaken for the purpose of ascertaining which one of a vocabulary of words (or, more generally, utterances) an input voice signal most clearly resembles, and information is available as to the a priori probabilities associated with the words of the vocabulary. One example of such a situation is the automatic telephone directory enquiry system described in our co-pending International patent application No. WO95/02524. In that system, several towns as having the closest matches to the spoken town name, and produces a "score" or probability indicating the closeness of the match; towns; list, having the closest matches to the spoken road name, again with scores; for the town the road is located in, and the most likely "road" result considered to be the one with the best weighted score.
The a priori probabilities do not necessarily have to originate from a preceding speech recognition process; for example another directory enquiry system, also described in the above-noted patent application, uses signals identifying the origin of a call to access statistical information as to the most likely towns to be wanted by an enquirer from that area to weight the results of a town name recognition process.
This process has the advantage of reliability--the retention of, for example, the second choice towns does not result in the selection of a road in that town unless that road scores markedly better in the road name recognition step than the first choice town. A disadvantage of this process however is that because the recogniser, when performing the road-name recognition step, produces only a limited number of candidate road names, it is possible that this short-list of road names may contain only the names of roads located in the more poorly-scoring towns--i.e. poorly-scoring road names, of roads located in high-scoring towns, have already been "pruned out" by the recogniser before the weighting process can be applied.
U.S. Pat. No. 4,783,803 describes a speech recognition apparatus in which the a priori probabilities relate to the given context of one or more patterns that have previously been recognised. A language score indicative of the probability of a certain word occurring after another certain word is combined with the score obtained for a sequence containing those words.


SUMMARY OF THE INVENTION

According to the present invention there is provided a method of speech recognition comprising: models to generate, for each of a plurality of allowable sequences of reference utterances defined by stored data defining such sequences, accumulated measures of similarity including contributions from previously generated measures obtained from comparison of one or more earlier portions of the utterance with a reference model or models corresponding to an earlier utterance or utterances in the respective allowable sequence, excluding from further repetitive comparison any sequence for which the accumulated measure is, to a degree defined by a predetermined pruning criterion, less indicative of similarity than the measures for other such sequences, and weighting the accumulated measures in accordance with weighting factors for each of the allowed sequences wherein the weighting is performed by weighting each computation of a measure or accumulated measure for a partial sequence by combined values of the weighting factors for each of the allowable sequences which commence with that partial sequence, less any such weighting factors applied to a measure generated in respect of an utterance or shorter sequence with which that partial sequence commences.
Preferably any sequence for which the weighted accumulated measure is, to a degree defined by a pruning criterion, less indicative of similarity than the measures for other such sequences is excluded from furthe

REFERENCES:
patent: 4783803 (1988-11-01), Baker et al.
patent: 5202952 (1993-04-01), Gillick et al.
patent: 5749069 (1998-05-01), Komori et al.
Ali et al, "Syllable Recognition Using an Adaptive Pattern Recognition Technique", Nov. 25, 1992, Journal of The Institution of Electronics & Telecommunication Engineers, pp. 677-683.

Affiliated with

Also associated with

Rating

Speech recognition incorporating a priori probability weighting does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Speech recognition incorporating a priori probability weighting , we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Speech recognition incorporating a priori probability weighting will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-835743