Distributed voice recognition system using acoustic feature vector modification

• US 20020103639A1
• Filed: 01/31/2001
• Published: 08/01/2002
• Est. Priority Date: 01/31/2001
• Status: Active Grant

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Chang, Chienchung, Malayath, Naren, Yafuso, Byron Yoshio
• Granted Patent: US 7,024,359 B2
• Time in Patent Office: Days
• US Class Current: 704/223
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications

Abstract

A voice recognition system applies speaker-dependent modification functions to acoustic feature vectors prior to voice recognition pattern matching against a speaker-independent acoustic model. An adaptation engine matches a set of acoustic feature vectors X with an adaptation model to select a speaker-dependent feature vector modification function f( ), which is then applied to X to form a modified set of acoustic feature vectors f(X). Voice recognition is then performed by correlating the modified acoustic feature vectors f(X) with a speaker-independent acoustic model.

Claims

1. A voice recognition system comprising:

an adaptation model containing acoustic pattern information; and

an adaptation engine for performing pattern matching of acoustic feature vectors with the acoustic pattern information to identify a selected feature vector modification function.

2. The voice recognition system of claim 1 wherein the adaptation engine is further configured to apply the selected feature vector modification function to the acoustic feature vectors to produce a set of modified acoustic feature vectors.

3. The voice recognition system of claim 1 further comprising a voice recognition engine for matching the set of modified acoustic feature vectors with an acoustic model.

4. The voice recognition system of claim 1 further comprising a control processor for evaluating the performance of the selected feature vector modification function and adjusting the selected feature vector modification function based on the evaluating.

5. The voice recognition system of claim 1 further comprising a memory for storing at least one set of parameters corresponding to a set of feature vector modification functions, wherein the selected feature vector modification function is a member of the set of feature modification functions.

6. The voice recognition system of claim 5 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a specific speaker.

7. The voice recognition system of claim 5 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a different acoustic environment.

8. A voice recognition system comprising:

an adaptation model containing acoustic pattern information; and

an adaptation engine for performing pattern matching of acoustic feature vectors with the acoustic pattern information to identify a speaker-dependent feature vector modification function and apply the speaker-dependent feature vector modification function to the acoustic feature vectors.

9. A voice recognition system comprising:

an adaptation engine for analyzing acoustic feature vectors to identify a selected feature vector modification function and applying the selected feature vector modification function to the acoustic feature vectors to produce a set of modified acoustic feature vectors;

an acoustic model; and

a voice recognition engine for matching the set of modified acoustic feature vectors with the acoustic model.

10. The voice recognition system of claim 9 further comprising an adaptation model, wherein said adaptation engine performs the analyzing acoustic feature vectors by matching the acoustic feature vectors with acoustic patterns stored in the adaptation model.

11. The voice recognition system of claim 9 further comprising a control processor for evaluating the performance of the selected feature vector modification function and adjusting the selected feature vector modification function based on the evaluating.

12. The voice recognition system of claim 9 further comprising a memory for storing at least one set of parameters corresponding to a set of feature vector modification functions, wherein the selected feature vector modification function is a member of the set of feature modification functions.

13. The voice recognition system of claim 12 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a specific speaker.

14. The voice recognition system of claim 12 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a different acoustic environment.

16. The remote station apparatus of claim 15 further comprising a control processor for evaluating the performance of the selected feature vector modification function and adjusting the selected feature vector modification function based on the evaluating.

17. The remote station apparatus of claim 15 further comprising a memory for storing at least one set of parameters corresponding to a set of feature vector modification functions, wherein the selected feature vector modification function is a member of the set of feature modification functions.

18. The remote station apparatus of claim 17 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a specific speaker.

19. The remote station apparatus of claim 17 wherein the memory contains more than one set of parameters corresponding to a set of feature vector modification functions, and wherein each set of parameters corresponds to a different acoustic environment.

20. The remote station apparatus of claim 15 further comprising a communications interface for communicating the modified acoustic feature vectors to a communications center.

15. A remote station apparatus comprising:

an adaptation model containing acoustic pattern information; and

an adaptation engine for performing pattern matching of acoustic feature vectors against the acoustic pattern information to identify a selected feature vector modification function, and applying the selected feature vector modification function to the acoustic feature vectors to produce a set of modified acoustic feature vectors.

21. A voice recognition communication center apparatus comprising:

a communications interface for receiving acoustic feature vectors, wherein the acoustic feature vectors have been modified using a feature vector modification function;

an acoustic model containing acoustic patterns, wherein the acoustic model is not trained to a single speaker;

a voice recognition engine for matching the set of modified acoustic feature vectors with the acoustic model; and

a control processor for evaluating the feature modification function based on the matching.

22. A voice recognition communication center apparatus comprising:

a communications interface for receiving acoustic feature vectors and speaker identification information from at least one remote station;

a memory containing speaker-dependent feature vector modification function parameters; and

an adaptation engine for performing pattern matching of the acoustic feature vectors with an acoustic model, identifying a selected speaker-dependent feature vector modification function based on the pattern matching and the speaker identification information, and applying the selected speaker-dependent feature vector modification function to the acoustic feature vectors to produce a set of modified acoustic feature vectors.

23. The voice recognition communication center apparatus of claim 22 wherein the adaptation engine further comprises a voice recognition engine for matching the set of modified acoustic feature vectors with a central acoustic model, wherein the central acoustic model is not trained to a single speaker.

24. The voice recognition communication center apparatus of claim 22 further comprising a central acoustic model, wherein the central acoustic model is not trained to a single speaker, and wherein the adaptation engine is further configured to perform pattern matching of the set of modified acoustic feature vectors with the central acoustic model.

25. The voice recognition communication center apparatus of claim 22 further comprising a voice recognition engine for matching the set of modified acoustic feature vectors with a central acoustic model.

26. The voice recognition communication center apparatus of claim 22 further comprising a control processor for evaluating the performance of the selected speaker-dependent feature vector modification function and adjusting the parameters of the selected feature speaker-dependent vector modification function in the memory based on the evaluating.

28. The method of claim 27 wherein the feature vector modification function is selected from a speaker-dependent set of feature vector modification functions.

29. The method of claim 28 further comprising modifying the speaker-dependent set of feature vector modification functions based on the adaptation pattern matching.

30. The method of claim 28 further comprising modifying the speaker-dependent set of feature vector modification functions based on the voice recognition pattern matching.

31. The method of claim 27 wherein the feature vector modification function is specific to an acoustic environment.

27. A method of performing voice recognition comprising:

extracting acoustic feature vectors;

performing adaptation pattern matching of the acoustic feature vectors with an adaptation model;

selecting a feature vector modification function based on the adaptation pattern matching;

applying the selected feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors; and

performing voice recognition pattern matching of the set of modified acoustic feature vectors with an acoustic model.

32. A method of performing voice recognition comprising:

at a remote station, performing adaptation pattern matching of acoustic feature vectors with an adaptation model stored in the remote station;

at the remote station, selecting a feature vector modification function from feature vector modification function information stored at the remote station, wherein the selecting a feature vector modification function is based on the pattern matching;

at the remote station, applying the selected feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors; and

sending the modified acoustic feature vectors from the remote station to a communications center.

33. The method of claim 32 wherein the feature vector modification function information is speaker-dependent.

34. The method of claim 32 further comprising modifying the feature vector modification function information based on the adaptation pattern matching.

35. The method of claim 32 further comprising modifying the feature vector modification function information at the remote station based on information received from the communications center.

36. The method of claim 32 further comprising performing voice recognition at the remote station, wherein the performing voice recognition comprises performing voice recognition pattern matching of the modified acoustic feature vectors with an acoustic model stored in the remote station.

37. The method of claim 36 further comprising modifying the feature vector modification function information at the remote station based on the voice recognition pattern matching.

38. The method of claim 32 wherein the feature vector modification function information is environment-dependent.

40. The method of claim 39 further comprising, at the communications center, selecting a speaker-dependent set of feature vector modification functions, wherein the selected feature vector modification function is selected from the speaker-dependent set of feature vector modification functions.

41. The method of claim 40 further comprising modifying the speaker-dependent set of feature vector modification functions based on the adaptation pattern matching.

42. The method of claim 40 further comprising modifying the speaker-dependent set of feature vector modification functions based on the voice recognition pattern matching.

43. The method of claim 40 further comprising sending, from the remote station to the communications center, speaker identification information, wherein the selecting a speaker-dependent set of feature vector modification functions is based on the speaker identification information.

39. A method of performing voice recognition in a system comprising at least one remote station and a communications center, the method comprising:

at the remote station, extracting acoustic feature vectors;

sending the acoustic feature vectors from the remote station to the communications center;

at the communications center, performing adaptation pattern matching of the acoustic feature vectors with an adaptation model stored in the communications center;

at the communications center, selecting a feature vector modification function from a database stored in the communications center, wherein the selecting a feature vector modification function is based on the adaptation pattern matching;

at the communications center, applying the selected feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors; and

at the communications center, performing voice recognition pattern matching of the set of modified acoustic feature vectors with an acoustic model stored in the communications center.

44. A method of performing voice recognition in a system comprising at least one remote station and a communications center, the method comprising:

at the remote station, extracting unmodified acoustic feature vectors;

at the remote station, performing adaptation pattern matching of the unmodified acoustic feature vectors with an adaptation model stored in the remote station;

at the remote station, selecting a speaker-dependent feature vector modification function based on the adaptation pattern matching;

at the remote station, applying the selected speaker-dependent feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors;

sending the modified acoustic feature vectors from the remote station to the communications center;

at the communications center, performing voice recognition pattern matching of the set of modified acoustic feature vectors with an acoustic model stored in the communications center.

45. The method of claim 44 further comprising modifying the selected speaker-dependent feature vector modification function based on the adaptation pattern matching.

46. The method of claim 44 further comprising:

sending the unmodified acoustic feature vectors from the remote station to the communications center;

at the communications center, analyzing the selected speaker-dependent feature vector modification function using the modified acoustic feature vectors and the unmodified acoustic feature vectors; and

at the remote station, modifying the selected speaker-dependent feature vector modification function based on the analyzing.

47. The method of claim 44 further comprising modifying the speaker-dependent set of feature vector modification functions based on the voice recognition pattern matching.

48. A voice recognition system comprising:

means for extracting acoustic feature vectors;

means for performing adaptation pattern matching of the acoustic feature vectors with an adaptation model;

means for selecting a feature vector modification function based on the adaptation pattern matching;

means for applying the selected feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors; and

means for performing voice recognition pattern matching of the set of modified acoustic feature vectors with an acoustic model.

49. A remote station apparatus comprising:

means for performing adaptation pattern matching of acoustic feature vectors with an adaptation model stored in the remote station;

means for selecting a feature vector modification function from feature vector modification function information stored at the remote station, wherein the selecting a feature vector modification function is based on the pattern matching;

means for applying the selected feature vector modification function to the acoustic feature vectors to form a set of modified acoustic feature vectors; and

means for sending the modified acoustic feature vectors to a communications center.