Word-dependent language model

• US 8,838,449 B2
• Filed: 12/23/2010
• Issued: 09/16/2014
• Est. Priority Date: 12/23/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Ju, Yun-Cheng, Tashev, Ivan J., Heinemann, Chad R.
• Primary Examiner(s): Wozniak, James
• Assistant Examiner(s): Nguyen, Timothy
• Application Number:
  US12/977,461
• Publication Number:
  US 20120166196A1
• Time in Patent Office: 1,363 Days
• Field of Search: 704/254
• US Class Current: 704/254
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

This document describes word-dependent language models, as well as their creation and use. A word-dependent language model can permit a speech-recognition engine to accurately verify that a speech utterance matches a multi-word phrase. This is useful in many contexts, including those where one or more letters of the expected phrase are known to the speaker.

Claims

1. A computer-implemented method comprising:

determining, based on a multi-word phrase, a word-dependent language model specific to the multi-word phrase and enabling word-level speech verification, the word-dependent language model including different accuracy thresholds for at least two of the multiple words, the different accuracy thresholds including a highest accuracy threshold for a most important word in the multi-word phrase, the highest accuracy threshold generated by providing a greatest number of imposter words in the word-dependent language model for the most important word in the multi-word phrase, wherein determining the word-dependent language model determines, based on a length of the most important word in the multi-word phrase, a set of words having the length, and wherein the imposter words for the most important word in the word-dependent language model includes the set of words;

providing the word-dependent language model of the multi-word phrase to a speech-recognition engine; and

receiving, from the speech-recognition engine and responsive to the speech-recognition engine receiving one or more human utterances associated with the multi-word phrase, an indication that the one or more utterances is verified or not verified to be the multi-word phrase.

2. A computer-implemented method as described in claim 1, wherein the set of words can be used as part of an in-vocabulary recognition algorithm for a portion of the one or more human utterances associated with the word of the multi-word phrase.

3. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines, based on one or more letters and letter locations of multiple words in the multi-word phrase that are known to a person from which the one or more human utterances are expected, the different accuracy thresholds for the at least two of the multiple words.

4. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines, based on one or more letters and letter locations of a word in the multi-word phrase that are known to a person from which the one or more human utterances are expected, a set of words having the one or more letters, and further comprising including the set of words in the word-dependent language model.

5. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines, based on a complete word of the multi-word phrase having been revealed to a person from which the one or more human utterances are expected, that the complete word does not need to be verified and further comprising including an indication that a portion of the one or more human utterances associated with this complete word does not need to be verified in the word-dependent language model.

6. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines higher-probability wrong words based on a length of a word in the multi-word phrase, and further comprising including the higher-probability wrong words in the word-dependent language model.

7. A computer-implemented method as described in claim 6, wherein the higher-probability wrong words are intended to be used as part of an out-of-vocabulary recognition algorithm for a portion of the one or more human utterances associated with the word of the multi-word phrase.

8. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines higher-probability wrong words based on a history of prior human utterances associated with a word in the multi-word phrase, and further comprising including the higher-probability wrong words in the word-dependent language model.

9. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines higher-probability wrong words based on common usage in a language in which the one or more human utterances are expected, and further comprising including the higher-probability wrong words in the word-dependent language model.

10. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model determines higher-probability wrong words for a word in the multi-word phrase based on one or more other words of the multi-word phrase that are adjacent to the word in the multi-word phrase, and further comprising including the higher-probability wrong words in the word-dependent language model.

11. A computer-implemented method as described in claim 1, wherein determining the word-dependent language model includes assigning different weights to different branches, each of the different branches associated with different words of the multi-word phrase and having one or more imposter words for respective different words of the multi-word phrase.

12. A computer-implemented method as described in claim 1, further comprising receiving the multi-word phrase from an entity and, responsive to receiving the indication that the one or more utterances is verified or not verified, passing the indication to the entity.

13. A system comprising:

one or more processors;

one or more memories comprising instructions stored thereon that, responsive to execution by the one or more processors, perform operations comprising;

determining, based on a multi-word phrase, a word-dependent language model specific to the multi-word phrase and enabling word-level speech verification, the word-dependent language model including different accuracy thresholds for at least two of the multiple words, the different accuracy thresholds including a highest accuracy threshold for a most important word in the multi word phrase, the highest accuracy threshold generated by providing a greatest number of imposter words in the word-dependent language model for the most important word in the multi-word phrase, wherein determining the word-dependent language model determines, based on a length of the most important word in the multi-word phrase, a set of words having the length, and wherein the imposter words for the most important word in the word-dependent language model includes the set of words;

providing the word-dependent language model of the multi-word phrase to a speech-recognition engine; and

receiving, from the speech-recognition engine and responsive to the speech recognition engine receiving one or more human utterances associated with the multi-word phrase, an indication that the one or more utterances is verified or not verified to be the multi-word phrase.

14. The system as described in claim 13, wherein the set of words can be used as part of an in-vocabulary recognition algorithm for a portion of the one or more human utterances associated with the word of the multi-word phrase.

15. The system as described in claim 13, wherein determining the word-dependent language model determines, based on one or more letters and letter locations of multiple words in the multi word phrase that are known to a person from which the one or more human utterances are expected, the different accuracy thresholds for the at least two of the multiple words.

16. The system as described in claim 13, wherein determining the word-dependent language model determines, based on one or more letters and letter locations of a word in the multi-word phrase that are known to a person from which the one or more human utterances are expected, a set of words having the one or more letters, and further comprising including the set of words in the word-dependent language model.

17. The system as described in claim 13, wherein determining the word-dependent language model determines, based on a complete word of the multi-word phrase having been revealed to a person from which the one or more human utterances are expected, that the complete word does not need to be verified and further comprising including an indication that a portion of the one or more human utterances associated with this complete word does not need to be verified in the word-dependent language model.

18. The system as described in claim 13, wherein determining the word-dependent language model determines higher probability wrong words based on a length of a word in the multi-word phrase, and further comprising including the higher-probability wrong words in the word-dependent language model.