Biasing a speech recognizer based on prompt context

• US 7,542,907 B2
• Filed: 12/19/2003
• Issued: 06/02/2009
• Est. Priority Date: 12/19/2003
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Lewis, James R., Epstein, Mark E.
• Primary Examiner(s): Smits; Talivaldis I
• Application Number:
  US10/741,465
• Publication Number:
  US 20050137868A1
• Time in Patent Office: 1,992 Days
• Field of Search: 704/251, 704/255, 704/257, 704/275
• US Class Current: 704/275
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models

Abstract

A method, a system, and an apparatus biasing a speech recognizer based on prompt context. The present invention is capable of analyzing the words used in the prompt given to the user. Then, a set of words the user is likely to say in response to the prompt is determined. The word set may be determined using a technology used by the speech recognition system, such as n-grams, grammars, or both. The speech recognition system boosts the probabilities of the analyzed words in the word set by a preconfigured amount. The preconfigured amount is selected based on collected data.

Claims

1. A method of biasing a speech recognition system at a lexical level based on prompt context comprising:

prompting a user for a response

analyzing words used in the prompt given to the user;

determining a set of words the user is likely to use when responding to the prompt based on the analyzing step, wherein the word set is determined using a speech recognition system strategy including at least one of n-gram language models and grammars; and

boosting probabilities of the determined set of words in a response from the user to the prompt by an adjustable predetermined amount such that a speech recognizer of the speech recognition system has an increased likelihood of recognizing the response from the user.

2. The method of claim 1, wherein the step of determining a response set comprises:

inverting question-asking syntax into a response syntax;

including at least one choice mentioned to the user; and

utilizing at least one additional constituent mentioned in the prompt.

3. The method of claim 1, wherein an interpolated grammar is formed by inducing a grammar from the response set and interpolating with a static grammar of what responses the user might say at any time.

4. The method of claim 3, wherein the interpolated grammar has weights of the responses and grammars adjusted therein based upon previous responses to the prompt.

5. The method of claim 3, wherein the interpolated grammar is used for future responses.

6. The method of claim 3, wherein the interpolated grammar is reset to the static grammar after the user responds to the prompt.

7. The method of claim 3, wherein the interpolated grammar is reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.

8. The method of claim 1 wherein a probabilistically adjusted grammar is formed by adapting a static grammar modeling of responses the user might say at any time to the phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar.

9. The method of claim 8, wherein the probabilistically adjusted grammar is reset to the static grammar after the user responds to the prompt.

10. The method of claim 8, wherein the probabilistically adjusted grammar is reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.

11. The method of claim 1 wherein an interpolated n-gram language model is formed by inducing a language model from the response set and interpolating with a static n-gram language model of responses the user might say at any time.

12. The method of claim 11, wherein the interpolated n-gram language model has weights of the induced language model and static language model adjusted therein based upon previous responses to the prompt.

13. The method of claim 11, wherein the interpolated n-gram language model is used for future responses.

14. The method of claim 13, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

15. The method of claim 13, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

16. The method of claim 1, wherein an adapted n-gram language model is formed by adapting a static n-gram language model of responses the user might say at any time with phrases in the response set to boost the probability of predicting phrases in the response set by the static n-gram language model.

17. The method of claim 16, wherein the adapted n-gram language model has weights of the adapted language model and static language model adjusted therein based upon previous responses to the prompt.

18. The method of claim 16, wherein the adapted n-gram language model is used for future responses.

19. The method of claim 18, wherein the adapted n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

20. The method of claim 18, wherein the adapted n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

21. The method of claim 1, wherein the speech recognition system strategy uses n-gram language models with embedded grammars for named entries.

22. The method of claim 21, wherein the n-gram language models with embedded grammars strategy comprises:

creating an interpolated n-gram language model that is formed by interpolating the response set and with a static n-gram language model of responses the user might say at any time;

analyzing responses in the response set to determine values mentioned for the embedded grammars; and

creating new embedded grammars from values mentioned in the prompt using one additional step;

wherein the additional step is selected from;

a) inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars; and

b) adapting static embedded grammars modeling of responses the user might say at any time is with phases in the response set to boost the probability of predicting phases in the response set by the static grammar to form probabilistically adjusted embedded grammars.

23. The method of claim 22, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the interpolated n-gram language model has weights of the responses adjusted therein based upon previous responses to the prompt.

24. The method of claim 23, wherein the interpolated n-gram language model is used for future responses.

25. The method of claim 24, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

26. The method of claim 24, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

27. The method of claim 23, wherein the interpolated embedded grammars have weights of the responses adjusted therein based upon previous responses to the prompt.

28. The method of claim 23, wherein the interpolated embedded grammars are used for future responses.

29. The method of claim 23, wherein the interpolated embedded grammars are reset to the static grammar after the user responds to the prompt.

30. The method of claim 23, wherein the interpolated embedded grammar are reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.

31. The method of claim 22, wherein the additional step is adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars;

further wherein the interpolated n-gram language model has weights of the responses adjusted therein based upon previous responses to the prompt.

32. The method of claim 31, wherein the interpolated n-gram language model is used for future responses.

33. The method of claim 32, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

34. The method of claim 32, wherein the interpolated n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

35. The method of claim 31, wherein the probabilistically adjusted embedded grammars are reset to the static grammar after the user responds to the prompt.

36. The method of claim 31, wherein the probabilistically adjusted embedded grammars are reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.

37. The method of claim 21, wherein the n-gram language models with embedded grammars strategy comprises:

creating an adapted n-gram language model that is formed by adapting a static n-gram language model of responses the user might say at any time is adapted with phrases in the response set to boost the probability of predicting phrases in the response set by the static n-gram language model;

analyzing responses in the response set to determine values mentioned for the embedded grammars; and

creating new embedded grammars from values mentioned in the prompt using one additional step;

wherein the additional step is selected from;

a) inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars; and

b) adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars.

38. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the adapted n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

39. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the adapted n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

40. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the interpolated embedded grammars have weights of the responses adjusted therein based upon previous responses to the prompt.

41. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the interpolated embedded grammars are used for future responses.

42. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the interpolated embedded grammars are reset to the static grammar after the user responds to the prompt.

43. The method of claim 37, wherein the additional step is inducing new embedded grammars from the response set and interpolating with a static grammar of what responses the user might say at any time to form interpolated embedded grammars;

further wherein the interpolated embedded grammars are reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.

44. The method of claim 37, wherein the additional step is adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars;

further wherein the adapted n-gram language model is reset to the static n-gram language model after the user responds to the prompt.

45. The method of claim 37, wherein the additional step is adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars;

further wherein the adapted n-gram language model is reset to the static n-gram language model after the user cancels a prompt and resets the system to generate a new prompt.

46. The method of claim 37, wherein the additional step is adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars;

further wherein the probabilistically adjusted embedded grammars are reset to the static grammar after the user responds to the prompt.

47. The method of claim 37, wherein the additional step is adapting static embedded grammars modeling of responses the user might say at any time is with phrases in the response set to boost the probability of predicting phrases in the response set by the static grammar to form probabilistically adjusted embedded grammars;

further wherein the probabilistic ally adjusted embedded grammars are reset to the static grammar after the user cancels a prompt and resets the system to generate a new prompt.