System and method of mu-law or A-law compression of bark amplitudes for speech recognition

• US 6,694,294 B1
• Filed: 10/31/2000
• Issued: 02/17/2004
• Est. Priority Date: 10/31/2000
• Status: Active Grant

• Current Assignee: Qualcomm, Inc.
• Original Assignee: Qualcomm, Inc.
• Inventors: Garudadri, Harinath
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): Storm, Donald L.
• Application Number:
  US09/703,191
• Time in Patent Office: 1,204 Days
• Field of Search: 704/234, 704/200.1, 704/211, 704/221, 704/226, 704/227
• US Class Current: 704/234
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence

Abstract

A method and system that improves voice recognition by improving the voice recognizer of a voice recognition system. Mu-law compression of bark amplitudes is used to reduce the effect of additive noise and thus improve the accuracy of the voice recognition system. A-law compression of bark amplitudes is used to improve the accuracy of the voice recognizer. Both mu-law compression and mu-law expansion can be used in the voice recognizer to improve the accuracy of the voice recognizer. Both A-law compression and A-law expansion can be used in the voice recognizer to improve the accuracy of the voice recognizer.

Claims

1. A voice recognizer of a distributed voice recognition system, comprising:

a bark amplitude generation module configured to convert a digitized speech signal to bark amplitudes;

a mu-log compression module coupled to the bark amplitude generation module, the mu-log compression module configured to perform mu-log compression of the bark amplitudes;

a RASTA filtering module coupled to the mu-log compression module, the RASTA filtering module configured to RASTA filter the mu-log bark amplitudes; and

a cepstral transformation module coupled to the RASTA filtering module, the cepstral transformation module configured to generate j static cepstral coefficients and j dynamic cepstral coefficients.

2. The voice recognizer of claim 1 further comprising a backend configured to process the j static cepstral coefficients and j dynamic cepstral coefficients and produces a recognition hypothesis.

3. The voice recognizer of claim 1, wherein the mu-log compression is G.711 mu-log compression.

4. The voice recognizer of claim 1, wherein the bark amplitude generation module is configured to convert a digitized speech signal to k bark amplitudes once very T milliseconds.

5. The voice recognizer of claim 4, wherein the cepstral transformation module is configured to generate j static cepstral coefficients and j dynamic cepstral coefficients every T milliseconds.

6. The voice recognizer of claim 5, wherein T equals 10.

7. The voice recognizer of claim 4, wherein k equals 16.

8. A voice recognizer of a distributed voice recognition system, comprising:

a bark amplitude generation module configured to convert a digitized speech signal to bark amplitudes;

an A-log compression module coupled to the bark amplitude generation module, the A-log compression module configured to perform A-log compression of the bark amplitudes;

a RASTA filtering module coupled to the A-log compression module, the RASTA filtering module configured to RASTA filter the A-log bark amplitudes; and

a cepstral transformation module coupled to the RASTA filtering module, the cepstral transformation module configured to generate j static cepstral coefficients and j dynamic cepstral coefficients.

9. The voice recognizer of claim 8 further comprising a backend configured to process the j static cepstral coefficients and j dynamic cepstral coefficients and produces a recognition hypothesis.

10. The voice recognizer of claim 8, wherein the A-log compression is G. 711 A-log compression.

11. The voice recognizer of claim 8, wherein the bark amplitude generation module is configured to convert a digitized speech signal to k bark amplitudes once very T milliseconds.

12. The voice recognizer of claim 11, wherein the cepstral transformation module is configured to generate j static cepstral coefficients and j dynamic cepstral coefficients every T milliseconds.

13. The voice recognizer of claim 12, wherein T equals 10.

14. The voice recognizer of claim 11, wherein k equals 16.

15. A voice recognizer of a distributed voice recognition system, comprising:

a bark amplitude generation module configured to converts a digitized speech signal to bark amplitudes;

a mu-log compression module coupled to the bark amplitude generation module, the mu-log compression module configured to perform mu-log compression of the bark amplitudes;

a RASTA filtering module coupled to the mu-log compression module, the RASTA filtering module configured to RASTA filters the mu-log bark amplitudes; and

a mu-log expansion module coupled to the RASTA filtering module, the mu-log expansion module configured to perform mu-log expansion of the filtered mu-log bark amplitudes.

16. The voice recognizer of claim 15 further comprising a backend configured to process the expanded bark amplitudes and produces a recognition hypothesis.

17. The voice recognizer of claim 15, wherein the mu-log compression and expansion is G.711 mu-log compression and expansion.

18. The voice recognizer of claim 15, wherein the bark amplitude generation module is configured to convert a digitized speech signal to k bark amplitudes once very T milliseconds.

19. The voice recognizer of claim 18, wherein the mu-log expansion module is configured to expand the filtered mu-log bark amplitudes into k expanded bark amplitudes.

20. The voice recognizer of claim 19, wherein T equals 10.

21. The voice recognizer of claim 18, wherein k equals 16.

22. A voice recognizer of a distributed voice recognition system, comprising:

a bark amplitude generation module configured to convert a digitized speech signal to bark amplitudes;

an A-log compression module coupled to the bark amplitude generation module, the A-log compression module configured to perform A-log compression of the bark amplitudes;

a RASTA filtering module coupled to the A-log compression module, the RASTA filtering module configured to RASTA filter the A-log bark amplitudes; and

an A-log expansion module coupled to the RASTA filtering module, the A-log expansion module configured to perform A-log expansion of the filtered A-log bark amplitudes.

23. The voice recognizer of claim 22 further comprising a backend configured to process the expanded bark amplitudes and produces a recognition hypothesis.

24. The voice recognizer of claim 22, wherein the A-log compression and expansion is G.711 A-log compression and expansion.

25. The voice recognizer of claim 22, wherein the bark amplitude generation module is configured to convert a digitized speech signal to k bark amplitudes once very T milliseconds.

26. The voice recognizer of claim 25, wherein the A-log expansion module is configured to expand the filtered A-log bark amplitudes into k expanded bark amplitudes.

27. The voice recognizer of claim 25, wherein k equals 16.

28. The voice recognizer of claim 27, wherein T equals 10.

29. A method of voice recognizer processing for voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

mu-log compressing the bark amplitudes;

RASTA-filtering the mu-log bark amplitudes; and

transforming cepstrally the mu-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients.

30. The method of claim 29, wherein the mu-log compressing is G.711 mu-log compressing.

31. The method of claim 29, wherein the coverting includes converting the digitized speech signal to k bark amplitudes once very T milliseconds.

32. The method of claim 31, wherein the transforming includes transforming cepstrally the mu-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients every T milliseconds.

33. The method of claim 32, wherein T equals 10.

34. The method of claim 31, wherein k equals 16.

35. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

mu-log compressing the bark amplitudes;

RASTA-filtering the mu-log bark amplitudes;

transforming cepstrally the mu-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients; and

producing a recognition hypothesis based on the j static cepstral coefficients and j dynamic cepstral coefficients.

36. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

A-log compressing the bark amplitudes;

RASTA-filtering the A-log bark amplitudes; and

transforming cepstrally the A-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients.

37. The method of claim 36, wherein the A-log compressing is G.711 A-log compressing.

38. The method of claim 36, wherein the coverting includes converting the digitized speech signal to k bark amplitudes once very T milliseconds.

39. The method of claim 38, wherein the transforming includes transforming cepstrally the A-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients every T milliseconds.

40. The method of claim 38, wherein k equals 16.

41. The method of claim 39, wherein T equals 10.

42. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

A-log compressing the bark amplitudes;

RASTA-filtering the A-log bark amplitudes;

transforming cepstrally the A-log bark amplitudes to j static cepstral coefficients and j dynamic cepstral coefficients; and

producing a recognition hypothesis based on the j static cepstral coefficients and j dynamic cepstral coefficients.

43. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

mu-log compressing the bark amplitudes;

RASTA-filtering the mu-log bark amplitudes; and

mu-log expanding the filtered mu-log bark amplitudes.

44. The method of claim 43, wherein the mu-log compressing is G.711 mu-log compressing.

45. The method of claim 43, wherein the coverting includes converting the digitized speech signal to k bark amplitudes once very T milliseconds.

46. The method of claim 45, wherein k equals 16.

47. The method of claim 46, wherein T equals 10.

48. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

mu-log compressing the bark amplitudes;

RASTA-filtering the mu-log bark amplitudes; and

mu-log expanding the filtered mu-log bark amplitudes; and

producing a recognition hypothesis based on the expanded mu-log bark amplitudes.

49. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

A-log compressing the bark amplitudes;

RASTA-filtering the A-log bark amplitudes; and

A-log expanding the filtered A-log bark amplitudes.

50. The method of claim 49, wherein the A-log compressing is G.711 A-log compressing.

51. The method of claim 49, wherein the coverting includes converting the digitized speech signal to k bark amplitudes once very T milliseconds.

52. The method of claim 51, wherein k equals 16.

53. The method of claim 52, wherein T equals 10.

54. A method of voice recognition, comprising:

converting a digitized speech signal to bark amplitudes;

A-log compressing the bark amplitudes;

RASTA-filtering the A-log bark amplitudes; and

A-log expanding the filtered A-log bark amplitudes; and

producing a recognition hypothesis based on the expanded A-log bark amplitudes.