Methods and systems for optimizing model adaptation for a speech recognition system

• US 7,895,039 B2
• Filed: 03/21/2007
• Issued: 02/22/2011
• Est. Priority Date: 02/04/2005
• Status: Active Grant

• Current Assignee: Vocollect, Inc. (Honeywell International Inc.)
• Original Assignee: Vocollect, Inc. (Honeywell International Inc.)
• Inventors: Pike, Jeffrey, Braho, Keith
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US11/688,920
• Publication Number:
  US 20070192101A1
• Time in Patent Office: 1,434 Days
• Field of Search: 704/251
• US Class Current: 704/251
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/065 Adaptation

Abstract

A method for efficient use of resources of a speech recognition system includes determining a recognition rate, corresponding to either recognition of instances of a word or recognition of instances of various words among a set of words, and determining an accuracy range of the recognition rate. The method may further include adjusting adaptation of a model for the word or various models for the various words, based on a comparison of at least one value in the accuracy range with a recognition rate threshold. An apparatus for efficient use of resources of a speech recognition system includes a processor adapted to determine a recognition rate corresponding to either recognition of instances of a one word or recognition of instances of various words among a set of words, and an accuracy range of the recognition rate. The apparatus may further include a controller adapted to adjust adaptation of a model for the word or various models for the various words, based on a comparison of at least one value in the accuracy range with a recognition rate threshold.

Claims

1. A method for efficient use of resources of a speech recognition system, comprising:

using at least one processor, determining a recognition rate, corresponding to either recognition of instances of a word or recognition of instances of various words among a set of words;

determining an accuracy range of the recognition rate, corresponding to a desired confidence level; and

using a recognition rate threshold and the desired confidence level upon which to base model adaptation, wherein a model for the word is adapted or various models for the various words are adapted, based upon a comparison of at least one value in the accuracy range with the recognition rate threshold.

2. The method of claim 1, further comprising:

withholding adapting the model or the various models, until a predetermined number of observations has occurred.

3. The method of claim 1, wherein the accuracy range is based on a number of observations of the system.

4. The method of claim 2, wherein the number of observations needed to determine whether to adapt the model is inversely related to the difference between the recognition rate and the recognition rate threshold.

5. The method of claim 1, wherein determining the recognition rate comprises dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

6. The method of claim 5, wherein the errors or correct observations are estimated errors or estimated correct observations determined from evaluating system behavior.

7. The method of claim 1, wherein adapting the model or the various models further comprises withholding adapting until the at least one value in the accuracy range is equal to the recognition rate threshold, or withholding adapting until the at least one value in the accuracy range exceeds the recognition rate threshold if the recognition rate is an error rate, or withholding adapting until the at least one value in the accuracy range is below the recognition rate threshold if the recognition rate is an accuracy rate.

8. The method of claim 1, wherein the recognition rate threshold is a dynamic value.

9. The method of claim 1, wherein the recognition rate threshold is based on a background noise level.

10. The method of claim 1, wherein the recognition rate threshold is based on a signal-to-noise ratio.

11. The method of claim 1, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.

12. The method of claim 1, wherein the recognition rate threshold is based a recognition rate corresponding to a particular user or users of the system.

13. A method for efficient use of resources of a speech recognition system, comprising:

using a recognition rate threshold and a desired confidence level of a recognition rate upon which to base adaptation of a model;

using at least one processor, determining a recognition rate;

determining an accuracy range of the recognition rate corresponding to the desired confidence level, the accuracy range related to a number of observations;

using a relationship between the recognition rate threshold, the recognition rate and the number of observations, which balances the accuracy of the recognition rate with expediency in making a model adaptation decision; and

basing a decision for model adaptation on using the relationship.

14. The method of claim 13, further comprising:

withholding model adaptation until a predetermined number of observations has occurred.

15. The method of claim 13, wherein the accuracy range is inversely related to the number of observations.

16. The method of claim 13, wherein the number of observations needed to determine whether to adapt is inversely related to the difference between the recognition rate and the recognition rate threshold.

17. The method of claim 13, wherein determining the recognition rate comprises dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

18. The method of claim 17, wherein the errors or correct observations of the system are estimated errors or correct observations determined from evaluating system behavior.

19. The method of claim 13, wherein the decision to adapt the model comprises deciding whether to perform the model adaptation or withhold the model adaptation.

20. The method of claim 13, wherein the recognition rate threshold is a dynamic value.

21. The method of claim 13, wherein the recognition rate threshold is based on a background noise level.

22. The method of claim 13, wherein the recognition rate threshold is based on a signal-to-noise ratio.

23. The method of claim 13, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.

24. A method for efficient use of resources of a speech recognition system comprising:

using at least one processor, determining a recognition rate, corresponding to either recognition of instances of a word or recognition of instances of various words among a set of words;

determining an accuracy range of the recognition rate; and

adjusting adaptation, of a model for the word or various models for the various words, based on at least one value in the accuracy range being equal to a recognition rate threshold, or based on at least one value in the accuracy range exceeding a recognition rate threshold if the recognition rate is an error rate, or based on at least one value in the accuracy range being below the recognition rate threshold if the recognition rate is an accuracy rate.

25. The method of claim 24, further comprising:

withholding adjusting the adaptation, until a predetermined number of observations has occurred.

26. The method of claim 24, further comprising updating the recognition rate and the accuracy range of the recognition rate after a predetermined number of observations has occurred.

27. The method of claim 24, wherein the accuracy range is based on a number of observations of the system.

28. The method of claim 24, wherein determining the recognition rate comprises dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

29. The method of claim 28, wherein the errors or correct observations of the system are estimated errors or correct observations determined from evaluating system.

30. The method of claim 24, wherein adjusting the adaptation comprises adapting the model or the various models or withholding adapting the model or the various models.

31. The method of claim 24, wherein adjusting adaptation of the model comprises withholding adapting the model or the various models until the at least one value in the accuracy range is equal to a recognition rate threshold, or withholding adapting the model or the various models until the at least one value in the accuracy range exceeds a recognition rate threshold if the recognition rate is an error rate, or withholding adapting the model or the various models until the at least one value in the accuracy range is below a recognition rate threshold if the recognition rate is an accuracy rate.

32. The method of claim 24, wherein the recognition rate threshold is a dynamic value.

33. The method of claim 24, wherein the recognition rate threshold is based on a background noise level.

34. The method of claim 24, wherein the recognition rate threshold is based on a signal-to-noise ratio.

35. The method of claim 24, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.

36. A apparatus for efficient use of resources of a speech recognition system, comprising:

a processor configured to determine a recognition rate, corresponding to either recognition of instances of a word or recognition of instances of various words among a set of words, and an accuracy range of the recognition rate corresponding to a desired confidence level; and

a controller configured to use a recognition rate threshold and the desired confidence level upon which to base model adaptation, by adapting of a model for the word or various models for the various words, based upon a comparison of at least one value in the accuracy range with the recognition rate threshold.

37. The apparatus of claim 36, wherein the processor withholds adapting the model or the various models, until a predetermined number of observations has occurred.

38. The apparatus of claim 36, wherein the accuracy range is based on a number of observations of the system.

39. The apparatus of claim 36, wherein the number of observations needed to determine whether to adapt the model is inversely related to the difference between the recognition rate and the recognition rate threshold.

40. The apparatus of claim 36, wherein the processor determines the recognition rate by dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

41. The apparatus of claim 40, wherein the errors or correct observations of the system are estimated errors or correct observations determined from evaluating system behavior.

42. The apparatus of claim 36, wherein the controller controls adaptation of the model or the various models by withholding adapting the model or the various models until the at least one value in the accuracy range is equal to the recognition rate threshold, or withholding adapting the model or the various models until the at least one value in the accuracy range exceeds the recognition rate threshold if the recognition rate is an error rate, or withholding adapting the model or the various models until the at least one value in the accuracy range is below the recognition rate threshold if the recognition rate is an accuracy rate.

43. The apparatus of claim 36, wherein the recognition rate threshold is a dynamic value.

44. The apparatus of claim 36, wherein the recognition rate threshold is based on a background noise level.

45. The apparatus of claim 36, wherein the recognition rate threshold is based on a signal-to-noise ratio.

46. The apparatus of claim 36, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.

47. The apparatus of claim 36, wherein the recognition rate threshold is based on a recognition rate corresponding to a particular user or users of the system.

48. A apparatus for efficient use of resources of a speech recognition system, comprising:

a selector configured to use a recognition rate threshold and a desired confidence level of a recognition rate, upon which to base adaptation of a model;

a processor configured to determine a recognition rate and an accuracy range of the recognition rate corresponding to the desired confidence level, the accuracy range related to a number of observations; and

a controller configured to use a relationship between the recognition rate threshold, the recognition rate and the number of observations, which balances the accuracy of the recognition rate with expediency in making a model adaptation decision and base a decision for model adaptation on using the relationship.

49. The apparatus of claim 48, further comprising:

withholding model adaptation until a predetermined number of observations has occurred.

50. The apparatus of claim 48, wherein the accuracy range is inversely related to the number of observations.

51. The apparatus of claim 48, wherein the number of observations needed to determine whether to adapt is inversely related to the difference between the recognition rate and the recognition rate threshold.

52. The apparatus of claim 48, wherein the processor determines the recognition rate by dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

53. The apparatus of claim 52, wherein the errors or correct observations of the system are estimated errors or correct observations determined from evaluating system behavior.

54. The apparatus of claim 48, wherein the decision to adapt the model comprises deciding whether to perform the model adaptation or withhold the model adaptation.

55. The apparatus of claim 48, wherein the recognition rate threshold is a dynamic value.

56. The apparatus of claim 48, wherein the recognition rate threshold is based on a background noise level.

57. The apparatus of claim 48, wherein the recognition rate threshold is based on a signal-to-noise ratio.

58. The apparatus of claim 48, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.

59. A apparatus for efficient use of resources of a speech recognition system comprising:

a processor configured to determine a recognition rate, corresponding to either recognition of instances of a word or recognition of instances of various words among a set of words, and an accuracy range of the recognition rate; and

a controller configured to adjust adaptation of a model for the word or various models for the various words, based on a comparison of at least one value in the accuracy range with a recognition rate threshold.

60. The apparatus of claim 59, further comprising:

withholding adjusting the adaptation, until a predetermined number of observations has occurred.

61. The apparatus of claim 59, wherein the processor updates the recognition rate and the accuracy range of the recognition rate after a predetermined number of observations has occurred.

62. The apparatus of claim 59, wherein the accuracy range is based on a number of observations of the system.

63. The apparatus of claim 59, wherein the processor determines the recognition rate by dividing a number of errors made by the system by a number of observations of the system or dividing a number of correct observations of the system by a number of observations of the system.

64. The apparatus of claim 63, wherein the errors or correct observations of the system are estimated errors or correct observations determined from evaluating system behavior.

65. The apparatus of claim 59, wherein the controller adjusts the adaptation by permitting adapting the model or the various models or withholding adapting the model or the various models.

66. The apparatus of claim 59, wherein the controller adjusts the adaptation of the model by withholding adapting the model or the various models until the at least one value in the accuracy range is equal to or exceeds a recognition rate threshold if the recognition rate is an error rate, or withholding adapting the model or the various models until the at least one value in the accuracy range is equal to or is below a recognition rate threshold if the recognition rate is an accuracy rate.

67. The apparatus of claim 59, wherein the recognition rate threshold is a dynamic value.

68. The apparatus of claim 59, wherein the recognition rate threshold is based on a background noise level.

69. The apparatus of claim 59, wherein the recognition rate threshold is based on a signal-to-noise ratio.

70. The apparatus of claim 59, wherein the recognition rate threshold is based on a number of words in a vocabulary of the system.