Method of evaluating an utterance in a speech recognition system

• US 6,226,612 B1
• Filed: 01/30/1998
• Issued: 05/01/2001
• Est. Priority Date: 01/30/1998
• Status: Expired due to Term

• Current Assignee: Google Technology Holdings LLC (Alphabet Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Srenger, Edward, Kushner, William M., Meunier, Jeffrey A.
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US09/016,214
• Time in Patent Office: 1,187 Days
• Field of Search: 704/231, 704/251, 704/255, 704/256, 704/236, 704/239, 704/240, 704/242, 704/244
• US Class Current: 704/256.2
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/088 Word spotting

Abstract

The present invention provides a method of calculating, within the framework of a speaker dependent system, a standard filler, or garbage model, for the detection of out-of-vocabulary utterances. In particular, the method receives new training data in a speech recognition system (202); calculates statistical parameters for the new training data (204); calculates global statistical parameters based upon the statistical parameters for the new training data (206); and updates a garbage model based upon the global statistical parameters (208). This is carried out on-line while the user is enrolling the vocabulary. The garbage model described in this disclosure is preferably an average speaker model, representative of all the speech data enrolled by the user to date. Also, the garbage model is preferably obtained as a by-product of the vocabulary enrollment procedure and is similar in it characteristics and topology to all the other regular vocabulary HMMs.

Claims

1. A method of evaluating an utterance in a speech recognition system, said method comprising the steps of:

receiving new training data in said speech recognition system;

calculating statistical parameters for said new training data;

calculating global statistical parameters based upon said statistical parameters for said new training data; and

updating a garbage model based upon said global statistical parameters.

2. The method of evaluating an utterance of claim 1 wherein said step calculating statistical parameters for said new training data comprises calculating a mean and a covariance for said new training data.

3. The method of evaluating an utterance of claim 1 wherein said step of calculating global statistical parameters comprises calculating a global mean and a global covariance for said new training data.

4. The method of evaluating an utterance of claim 1 wherein said step of calculating global statistical parameters comprises calculating global statistical parameters based upon previous global statistical parameters and said statistical parameters for said new training data.

5. The method of evaluating an utterance of claim 1 wherein said step of calculating global statistical parameters comprises calculating global statistical parameters based upon a previous global mean and a previous global variance derived from a data base of trained data available before the start of a training procedure and said mean and said variance for said new data.

6. The method of evaluating an utterance of claim 1 wherein said step of updating a garbage model comprises generating a single state HMM.

7. The method of evaluating an utterance of claim 1 further comprising steps of receiving new recognition data and comparing a model for said new recognition data to a plurality of models previously stored in said speech recognition system.

8. The method of evaluating an utterance of claim 7 wherein said step of comparing a model for said new recognition data further comprises steps of searching data on a frame-by-frame basis and generating an updated score for each model for each frame of said new recognition data.

9. The method of evaluating an utterance of claim 8 further comprising steps of weighing said updated score by a scale factor when said garbage model is evaluated and comparing said updated score of said garbage model and updated scores for each model for each frame of said new recognition data to select which model is a best model as evaluated against all other active models.

10. The method of evaluating an utterance of claim 9 wherein said step of weighing said updated score comprises adjusting said updated score by a scale factor according to environmental conditions.

11. The method of evaluating an utterance of claim 10 wherein said step of weighing said updated score further comprises a step of sensing changes in said environmental conditions.

12. The method of evaluating an utterance of claim 9 further comprising a step of rejecting said utterance if said garbage model has been selected as the best model when evaluated against all other active models.

13. The method of evaluating an utterance of claim 9 wherein said step of weighing said updated score comprises weighing said updated score on a frame-by-frame basis.

14. The method of evaluating an utterance of claim 7 wherein said step of comparing a model for said new recognition data comprises searching data on a model-by-model basis and generating an updated score for each model of said new recognition data.

15. The method of evaluating an utterance of claim 14 further comprising steps of weighing said updated score by a scale factor when said garbage model is evaluated and comparing said updated score of said garbage model and updated scores for each model for each frame of said new recognition data to select which model is a best model as evaluated against all other active models.

16. The method of evaluating an utterance of claim 15 wherein said step of weighing said updated score comprises adjusting said updated score by a scale factor according to environmental conditions.

17. The method of evaluating an utterance of claim 16 wherein said step of weighing said updated score further comprises a step of sensing changes in said environmental conditions.

18. The method of evaluating an utterance of claim 15 further comprising a step of rejecting said utterance if said garbage model has been selected as the best model when evaluated against all other active models.

19. The method of evaluating an utterance of claim 15 wherein said step of weighing said score comprises weighing said updated score on a frame-by-frame basis.

20. A method of evaluating an utterance in a speaker-dependent speech recognition system for receiving data, said method comprising the steps of:

receiving user provided training data;

calculating statistical parameters for said new training data;

calculating a global mean and global variance based upon said mean and said covariance for said new training data; and

updating a garbage model based upon said global mean and said global covariance.

21. The method of evaluating an utterance of claim 20 wherein said step of calculating statistical parameters for said new training data comprises calculating a mean and a covariance for said new training data.

22. The method of evaluating an utterance of claim 20 wherein said step of calculating global statistical parameters comprises calculating a global mean and a global covariance for said new training data.

23. The method of evaluating an utterance of claim 22 wherein said step of calculating global statistical parameters comprises calculating global statistical parameters based upon previous global statistical parameters and said statistical parameters for said new training data.

24. The method of evaluating an utterance of claim 20 wherein said step of calculating global statistical parameters comprises calculating global statistical parameters based upon a previous global mean and a previous global variance received from a data base of trained data available before the start of a training procedure and said mean and said covariance for said new data.

25. The method of evaluating an utterance of claim 20 wherein said step of updating a garbage model comprises generating a single state HMM.

26. The method of evaluating an utterance of claim 20 further comprising steps of receiving new recognition data and comparing a model for said new recognition data to a plurality of models previously stored in said speech recognition system.

27. The method of evaluating an utterance of claim 26 wherein said step of comparing a model for said new recognition data further comprises steps of searching data on a frame-by-frame basis and generating an updated score for each model for each frame of said new recognition data.

28. The method of evaluating an utterance of claim 27 further comprising steps of weighing said updated score by a scale factor when said garbage model is evaluated and comparing said updated score of said garbage model and updated scores for each model for each frame of said new recognition data to select which model is a best model as evaluated against all other active models.

29. The method of evaluating an utterance of claim 28 wherein said step of weighing said updated score comprises adjusting said updated score by a scale factor according to environmental conditions.

30. The method of evaluating an utterance of claim 29 wherein said step of weighing said updated score further comprises a step of sensing changes in said environmental conditions.

31. The method of evaluating an utterance of claim 28 further comprising a step of rejecting said utterance if said garbage model has been selected as the best model when evaluated against all other active models.

32. The method of evaluating an utterance of claim 28 wherein said step of weighing said score comprises weighing said updated score on a frame-by-frame basis.

33. The method of evaluating an utterance of claim 26 wherein said step of comparing a model for said new recognition data comprises searching data on a model-by-model basis.

34. The method of evaluating an utterance of claim 33 further comprising steps of weighing said updated score by a scale factor when said garbage model is evaluated and comparing said updated score of said garbage model and updated scores for each model for each frame of said new recognition data to select which model is a best model as evaluated against all other active models.

35. The method of evaluating an utterance of claim 34 wherein said step of weighing said updated score comprises adjusting said updated score by a scale factor according to environmental conditions.

36. The method of evaluating an utterance of claim 35 wherein said step of weighing said updated score further comprises a step of sensing changes in said environmental conditions.

37. The method of evaluating an utterance of claim 34 further comprising a step of rejecting said utterance if said garbage model has been selected as the best model when evaluated against all other active models.

38. The method of evaluating an utterance of claim 34 wherein said step of weighing said score comprises weighing said updated score on a frame-by-frame basis.

39. A method of evaluating an utterance in a speaker-dependent speech recognition system for receiving data, said method comprising the steps of:

calculating a global mean and a global covariance based upon previously received training data;

receiving new training data;

calculating a mean and a covariance for said new training data;

updating said global mean and said global covariance based upon said mean and said covariance for said new training data; and

updating a garbage model based upon an updated global mean and an updated global covariance.

40. A method of evaluating an utterance in a speaker-dependent speech recognition system for receiving data, said method comprising the steps of:

calculating a global mean and a global covariance based upon previously received training data;

receiving new training data;

calculating a mean and a covariance for said new training data;

updating said global mean and said global covariance based upon said mean and said covariance for said new training data;

updating a single state HMM model for all data based upon an updated global mean and an updated global covariance;

receiving new data for operating said speaker-dependent speech recognition system;

comparing a model for said new data to a plurality of models previously stored in said speaker-dependent speech recognition system;

determining which of said plurality of models is a single state HMM model for all data;

multiplying said single state HMM model for all data by a penalty weight;

comparing said plurality of models as modified by said penalty weight to determine a best model from the list of active models; and

rejecting said utterance if a garbage model has been selected as the best model from the list of active models.