Speech recognition language models
First Claim
Patent Images
1. A method of speech recognition using multiple language models, the method comprising:
- receiving language model results for speech recognition candidates for an utterance from multiple language models;
combining the results from the multiple language models according to a combination expression to produce combined language model results, the combination expression applying different combination weights to the results received from the respective language models;
selecting one of the candidates selecting based on the combined language model results;
adjusting at least one of the combination expression weights based on the selected candidate; and
repeating the combining, selecting, and adjusting for a subsequent utterance.
8 Assignments
0 Petitions
Accused Products
Abstract
Language model results are combined according to a combination expression to produce combined language model results for a set of candidates. A candidate is selected and the combination expression is adjusted using language model results associated with the selected candidate.
278 Citations
28 Claims
-
1. A method of speech recognition using multiple language models, the method comprising:
-
receiving language model results for speech recognition candidates for an utterance from multiple language models; combining the results from the multiple language models according to a combination expression to produce combined language model results, the combination expression applying different combination weights to the results received from the respective language models; selecting one of the candidates selecting based on the combined language model results; adjusting at least one of the combination expression weights based on the selected candidate; and repeating the combining, selecting, and adjusting for a subsequent utterance. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19)
-
-
20. A method of generating language model results for use in speech recognition, the method comprising:
-
prompting a user to identify a source of text; retrieving text from the source of text identified by the user; building a topic language model from the retrieved text; producing combined language model results by combining language model results from the topic language model with language model results from one or more other language models according to a combination expression, the combination expression applying different combination weights to the results received from the respective language models; and adjusting at least one of the combination weights based on performance of the language models. - View Dependent Claims (21, 22, 23)
-
-
24. A computer program, residing on a computer readable medium, for a speech recognition system comprising a processor and an input device, the computer program comprising instructions performing speech recognition by causing the processor to perform the following operations:
-
receive language model results for speech recognition candidates for an utterance from multiple language models; combine the results from the multiple language models according to a combination expression to produce combined language model results, the combination expression applying different combination weights to the results received from the respective language models; evaluate the candidates using the combined language model results to select one of the candidates; adjust at least one of the combination expression weights based on the selected candidate; and repeat the combining, selecting, and adjusting for a subsequent utterance. - View Dependent Claims (25)
-
-
26. A computer program, residing on a computer readable medium, for a speech recognition system comprising a processor and an input device, the computer program comprising instructions for generating a topic language model for use in speech recognition by causing the processor to perform the following operations:
-
prompt a user to identify a source of text; retrieve text from the source of text identified by the user; build a topic language model from the retrieved text; produce combined language model results by combining language model results from the topic language model with language model results from one or more other language models according to a combination expression, the combination expression applying different combination weights to the results received from the respective language models; and adjust at least one of the combination weights based on performance of the language models. - View Dependent Claims (27)
-
-
28. A method of speech recognition using multiple language models, the method comprising:
-
receiving language model results for speech recognition candidates for an utterance from multiple language models, the language models including at least one language model formed from recently recognized words; combining the results from the multiple language models according to a combination expression to produce combined language model results, the combination expression being of the form;
space="preserve" listing-type="equation">P.sub.c (w.sub.a)=Σ
λ
.sub.x P.sub.x (w.sub.a)where Pc (wa) represents the combined language model score for word a, Px (wa) represents the score for word a for language model x, and λ
x represents a combination expression weight for language model x;selecting one of the candidates based on the combined language model results; adjusting at least one of the combination expression weights, λ
x, based on the selected candidate; andrepeating the combining, selecting, and adjusting for a subsequent utterance.
-
Specification