Adaptive noise suppressor

• US 4,658,426 A
• Filed: 10/10/1985
• Issued: 04/14/1987
• Est. Priority Date: 10/10/1985
• Status: Expired due to Fees

• Current Assignee: ANTIN, HAROLD 520 E., Mark Antin
• Original Assignee: Harold Antin, Mark Antin
• Inventors: Kenworthy, Gary, Lynn, Douglas, Chabries, Douglas M., Christiansen, Richard W.
• Primary Examiner(s): Rubinson, Gene Z.
• Assistant Examiner(s): Schroeder, L. C.
• Application Number:
  US06/785,999
• Time in Patent Office: 551 Days
• Field of Search: 381/47, 381/94, 381/93, 381/102
• US Class Current: 381/94.3
• CPC Class Codes: G10L 21/0208 Noise filtering
• CPC Class Codes: H03H 21/0027 filtering in the frequency domain
• CPC Class Codes: H04R 2225/43 Signal processing in hearing aids to enhance the speech intelligibility
• CPC Class Codes: H04R 25/453 electronically
• CPC Class Codes: H04R 3/02 for preventing acoustic reaction , i.e. acoustic oscillatory feedback specially adapted for hearing aids H04R25/453

Abstract

An adaptive noise suppressor for providing noise filtered signals. The noise suppression device employs a vector gain μ for the weights of the filter wherein the vector μ is selected for each frequency bin to be inversely proportional to the power spectrum. A projection operator is utilized to remove the effects of circular convolution to produce a linear convolution result wherein the weights are readjusted in a manner to minimize the difference between the input signal and the filter output signal, thereby minimizing the error signal to produce a noise suppressed signal in the filtered output. A frequency suppression device utilizes the same principles of the vector μ and projection operator, but the output is taken from the error output of the filter.

Claims

1. A noise suppression device for providing a noise filtered speech signal comprising:

means for delaying an input signal X by a fixed delay Δ

;

means for transforming the delayed signal into the frequency domain;

weighted means, having weights and having inputs connected to the transforming means, for adaptably filtering the delayed, transformed signal using a mean square algorithm in the frequency domain;

means for transforming the output of the weighted means into a time domain signal comprising the noise filtered speech signal Y;

means, having as inputs the signals X and Y for subtracting the signal Y from X, the output of the subtracting means being a noise error signal ε

;

means for transforming the noise error signal ε

into the frequency domain to produce a transformed error signals;

means for multiplying the transformed noise error signals by a vector gain μ

;

projection operator means receiving the output of the multiplying means for removing the effects of circular convolution, the output of the projection operator means being fed back to the weighted means to cause the weights of the weighted means to be readjusted in a manner to minimize the difference between the signal X and the signal Y, thereby minimizing the error signal and producing a noise suppressed signal in the filtered signal Y.

2. A noise suppression device as recited in claim 1, wherein Δ

is on the order of 1-3 ms.

3. A noise suppression device as recited in claim 2, wherein Δ

is on the order of 1 ms.

4. A noise suppression device as recited in claim 1, wherein the vector gain μ

is selected in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum, whereby spectrum fidelity is preserved.

5. A noise suppression device as recited in claim 1, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and noise data and said device further comprises means for zeroing the last half of the ε

data signals in the sampled group of data for providing zeroed ε

data signals to said noise error signal transforming means and wherein said transforming means receive said zeroed ε

data signals for transforming same into the frequency domain.

6. A noise suppression device as recited in claim 5 further comprising means for combining a second group of zeroed ε

data signals at the end of a first group of zeroed ε

data signals for producing a full group of ε

data signals for providing same to said noise error signal transforming means and wherein said transforming means receives said full group of ε

data signals for transforming same into the frequency domain.

7. A noise suppression device as recited in claim 5, wherein said projection operator means comprises means for zeroing the last half of the weights in the time domain.

8. A noise suppression device as recited in claim 6, wherein said projection operator means comprises means for zeroing the last half of the weights in the time domain.

9. A noise suppression device as recited in claim 4, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and noise data and said device further comprises means for zeroing the last half of the ε

data signals in the sampled group of data for providing zeroed ε

data signals to said noise error signal transforming means and wherein said transforming means receive said zeroed ε

data signals for transforming same into the frequency domain.

10. A noise suppression device as recited in claim 2, wherein the vector gain μ

is selected in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum.

11. A noise suppression device as recited in claim 3, wherein the vector gain μ

is selected in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum.

12. A noise suppression device as recited in claim 1, further comprising means for multiplying said weights by a leak factor ρ

so as to decrease the value of said weights in inverse proportion to the power in the frequency spectrum.

13. A noise suppression device as recited in claim 1, wherein a minimum value of μ

is set for each frequency bin.

14. A noise suppression device as recited in claim 1, wherein Δ

is on the order of less than a pitch period.

15. A method of noise suppression for providing a noise filtered signal comprising the steps of:

(a) delaying an input signal X by a fixed delay Δ

;

(b) transforming the delayed signal into the frequency domain;

(c) adaptably filtering the delayed, transformed signal using a plurality of weights and a mean square algorithm in the frequency domain to provide a filtered output signal;

(d) transforming the filtered output signals into a time domain signal comprising the noise filtered signal Y;

(e) subtracting the signal Y from X for obtaining a noise error signal ε

;

(f) transforming the noise error signal ε

into the frequency domain to produce a transformed error signals;

(g) multiplying the transformed noise error signals by a vector gain μ

;

(h) removing the effects of circular convolution from the multiplied noise error signal; and

(i) feeding back the results of step (h) to adjust the weights in the mean square algorithm in a manner to minimize the difference between the signal X and the signal Y, thereby minimizing the error signal ε and

producing noise suppressed signals in the noise filtered signal Y.

16. The method of claim 15, wherein the step of delaying the input signal comprises delaying the input signal on the order of 1-3 ms.

17. The method of claim 15, wherein the step of delaying the input signal comprises delaying the input signal on the order of 1 ms.

18. The method of claim 15 further comprising the step of selecting the vector gain μ

in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum, whereby spectrum fidelity is preserved.

19. The method of claim 15, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and noise data and, after said subtracting step (e), said method further comprises the step of zeroing the first half of the ε

data signals in the sampled group of data for providing zeroed ε

data signals for transforming in accordance with step (f).

20. The method of claim 19 further comprising the steps of:

combining a second group of zeroed ε

data signals at the end of a first group of zeroed ε

data signals for producing a full group of ε

data signals; and

providing said full group of ε

data signals for transforming same in accordance with step (f).

21. The method of claim 19, wherein step (h) comprises the step of zeroing the last half of the weights in the time domain.

22. The method of claim 20, wherein step (h) comprises the step of zeroing the last half of the weights in the time domain.

23. The method of claim 18, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and noise data and, after said subtracting step (e), said method further comprises the step of zeroing the first half of the ε

data signals in the sampled group of data for providing zeroed ε

data signals for transforming in accordance with step (f).

24. The method of claim 16 further comprising the step of selecting the vector gain μ

in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum, whereby spectrum fidelity is preserved.

25. The method of claim 17 further comprising the step of selecting the vector gain μ

in each frequency bin of the frequency spectrum to be inversely proportional to the power in the frequency spectrum, whereby spectrum fidelity is preserved.

26. The method of claim 15 further comprising, after step (f), the step of multiplying said weights by a leak factor ρ

so as to decrease the value of said weights in inverse proportion to the power in the frequency spectrum.

27. The method of claim 15 further comprising, after step (f), the step of selecting a minimum value of μ

for each frequency bin.

28. The method of claim 15, wherein the step of delaying the input signal comprises delaying the input signal on the order of less than a pitch period.

35. A feedback suppression device as recited in claim 28, wherein Δ

is about 100 ms.

53. The method of claim 27 further comprising, after step (f), the step of selecting a minimum value of μ

for each frequency bin.

29. A feedback suppression device for providing feedback suppressed signals comprising:

(a) means for delaying an input speech signal X by a fixed time delay Δ

;

(b) first means for transforming the delayed signal into the frequency domain;

(c) weighted means, having weights and having inputs connected to the first transforming means for adaptably filtering the delayed, transformed signal using a mean square algorithm in the frequency domain;

(d) second means for transforming the output of the weighted means into a time domain signal comprising a signal Y;

(e) means having as inputs the signals X and Y for subtracting the signal Y from X, the output of the subtracting means being a feedback suppressed signal ε

;

(f) third means for transforming the signal ε

into the frequency domain to produce a transformed feedback suppressed signals;

(g) means for multiplying the transformed feedback suppressed signal by a vector gain μ

; and

(h) projection operator means receiving the output of the multiplying means for removing the effects of circular convolution, the output of the projection operator means being fed back to the weighted means to cause the weights of the weighted means to be readjusted in a manner to minimize the difference between the signals X and Y, thereby minimizing the feedback suppressed signal ε

.

30. A feedback suppression device as recited in claim 29. wherein the number of weights of said weighted means is in the range of 16-32 weights.

31. A feedback suppression device as recited in claim 29, wherein the number of weights is about 22.

32. A feedback suppression device as recited in claim 29, wherein Δ

is on the order of 100-500 ms.

33. A feedback suppression device as recited in claim 29, wherein Δ

is about 100 ms.

34. A feedback suppression device as recited in claim 30, wherein Δ

is on the order of 50-500 ms.

36. A feedback suppression device as recited in claim 31, wherein Δ

is on the order of 50-500 ms.

37. A feedback suppression device as recited in claim 31, wherein Δ

is about 100 ms.

38. A feedback suppression device as recited in claim 29, wherein Δ

is on the order of 50-500 ms.

39. A feedback suppression device as recited in claim 29, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and feedback data and said device further comprises means for zeroing the first half of the ε

data signal in the sampled group of data for providing zeroed ε

data signals to said third transforming means and wherein said third transforming means receives said zeroed ε

data signals for transforming same into the frequency domain.

40. A feedback suppression device as recited in claim 39 further comprising means for combining a second group of zeroed ε

data signals at the end of a first group of zeroed ε

data signals for producing a full group of ε

data signals for providing same to said third transforming means and wherein said third transforming means receives said full group of ε

data signals for transforming same into the frequency domain.

41. A feedback suppression device as recited in claim 29, wherein said projection operator means comprises means for zeroing the last half of the weights in the time domain.

42. A feedback suppression device as recited in claim 40, wherein said projection operator means comprises means for zeroing the last half of the weights in the time domain.

43. A feedback suppression device as recited in claim 38, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and feedback data and said device further comprises means for zeroing the first half of the ε

data signal in the sampled group of data for providing zeroed ε

data signals to said third transforming means and wherein said third transforming means receives said zeroed ε

data signals for transforming same into the frequency domain.

44. A feedback suppression device as recited in claim 40, wherein Δ

is on the order of 50-500 ms.

45. A feedback suppression device as recited in claim 41, wherein Δ

is on the order of 50-500 ms.

46. A feedback suppression device as recited in claim 29 further comprising means for multiplying said weights by a leak factor ρ

so as to decrease the value of said weights in inverse proportion to the power in the frequency spectrum.

47. A feedback suppresion device as recited in claim 29, wherein Δ

is selected to be on the order of at least greater than a phonem.

48. A method of feedback suppression to provide feedback suppressed signals comprising the steps of:

(a) delaying an input signal X by a fixed time delay Δ

;

(b) transforming the delayed signal into the frequency domain;

(c) adaptably filtering the delayed, transformed signal using a plurality of weights and a mean square algorithm in the frequency domain to provide filtered output signals;

(d) transforming the filtered output signals of the weighted means into a time domain signal comprising a signal Y;

(e) subtracting the signal Y from X for obtaining a feedback suppressed signal ε

;

(f) transforming the signal ε

into the frequency domain to produce a transformed feedback suppressed signals;

(g) multiplying the transformed feedback suppressed signal by a vector gain μ

;

(h) removing the effects of circular convolution from the multiplied noise error signal; and

(i) feeding back the results of step (h) to adjust the weights in the mean square algorithm in a manner to minimize the difference between the signals X and Y, thereby minimizing the feedback suppressed signal ε

.

49. The method of claim 48, wherein the step of delaying comprises delaying the input signal on the order of at least greater than a phonem.

50. The method of claim 48, wherein the step of delaying the input signal comprises delaying the input signal on the order of 50-500 ms.

51. The method of claim 48, wherein said signals X, Y, and ε

comprise a plurality of digital signals representing a sampled group of signal and feedback data and, after step (e), said method further comprises the step of zeroing the first half of the ε

data signal in the sampled group for transforming in accordance with step (f).

52. The method of claim 48 further comprising, after step (f), the step of multiplying said weights by a leak factor ρ

so as to decrease the value of said weights in inverse proportion to the power in the frequency spectrum.

54. A time domain feedback suppression device comprising:

means for delaying a speech input signal X by a fixed delay Δ

, Δ

being selected to be on the order of at least greater than a phonem;

weighted means having weights and having inputs connected to the delaying means for adaptably filtering the delayed signal using a least-mean-square algorithm, the output of the filtering means comprising a signal Y;

means, having as inputs the signals X and Y for subtracting the signal Y from X, the output of the subtracting means being a feedback suppressed speech signal ε

;

means having as an input signal ε

for multiplying the signal by a gain μ

, the output of the multiplying means being fed back to the weighted means to cause the weights of the weighted means to be readjusted in a manner to minimize the difference between the signal X and the signal Y, thereby minimizing the feedback suppressed speech signal ε

.

55. A feedback suppression device as recited in claim 54, wherein the number of weights of said weighted means is in the range of 16-32 weights.

56. A feedback suppression device as recited in claim 55, wherein the number of weights is about 22.

57. A feedback suppression device as recited in claim 54, wherein Δ

is on the order of 50-500 ms.

58. A feedback suppression device as recited in claim 54, wherein Δ

is about 100 ms.

59. A feedback suppression device as recited in claim 55, wherein Δ

is on the order of 50-500 ms.

60. A feedback suppression device as recited in claim 54, wherein Δ

is about 100 ms.

61. A feedback suppression device as recited in claim 56, wherein Δ

is on the order of 50-500 ms.

62. A feedback suppression device as recited in claim 56, wherein Δ

is about 100 ms.

63. A method of feedback suppression comprising the steps of:

delaying a speech input signal X by a fixed delay Δ

, Δ

being selected to be on the order of at least a phonem;

adaptably filtering the delayed signal using a plurality of weights and a least-mean-square algorithm, the output of the filtering means comprising a signal Y;

subtracting the signal Y from X, for providing a feedback suppressed speech signal ε

;

multiplying the feedback suppressed speech signal ε

by a gain μ

; and

feeding back to the weighted means to cause the weights of the weighted means to be readjusted in a manner to minimize the difference between the signal X and the signal Y, thereby minimizing the feedback suppressed speech signal ε

.