Noise reduction system and method for audio device with multiple microphones

1. A method of voice processing for an audio device with an array of at least two microphones with dynamically changing array orientation to reduce background noise, comprising:

obtaining a multi-dimensional spatial feature vector comprising at least a correlation of the at least two microphones and a calculation of at least one ratio of energies of the at least two microphones;

using the multi-dimensional feature vector to estimate an energy of near-field speech and to estimate an energy of background noise;

using a ratio of the near-field speech energy estimate and the background noise energy estimate to estimate a probability of a presence of the near-field speech;

adaptively combining signals from the at least two microphones based on the estimated near-field speech presence probability to provide a combined output signal comprising a near-field speech signal and a residual background noise signal;

estimating a power spectral density of the residual background noise signal present at the combined output signal using the estimated near-field speech presence probability; and

reducing the background noise by using the estimated power spectral density.

2. The method of claim 1 , wherein said reducing the background noise comprises calculating a time-frequency dependent gain function based on the estimated residual background noise power spectral density and applying the calculated gain on the combined output signal to reduce the background noise.

3. The method of claim 1 , further comprising:

adaptively controlling said estimating the power spectral density of the residual background noise signal when at least one of the following varies: the background noise energy and the near-field speech energy.

4. The method of claim 3 , wherein a level of the background noise is discretized as N-state noise level modes and M-state SNR modes, wherein N and M are integers greater than one.

5. The method of claim 4 , further comprising:

dynamically modifying thresholds as a function of prevailing conditions of the background noise level and SNR; and

comparing the spatial features of the multi-dimensional feature vector with the dynamically modified thresholds to generate control signals.

6. The method of claim 1 , further comprising:

determining that the background noise comprises wind noise;

selecting as an output signal one of the following based on diversity of energy of the wind noise among the at least two microphones:

the combined output signal; and

the signal from one of the at least two microphones;

estimating a power spectral density of the wind noise present at the selected output signal; and

reducing the wind noise by using the estimated power spectral density of the wind noise.

7. The method of claim 6 , wherein the power spectral density of the wind noise is estimated using the multi-dimensional spatial feature vector.

8. The method of claim 7 , wherein the power spectral density of the wind noise is estimated differently depending on whether acoustic noise or wind noise is present.

9. The method of claim 1 , wherein the array of at least two microphones are spaced at least 100 millimeters apart.

10. The method of claim 1 , wherein the multi-dimensional spatial feature vector further comprises one or more spatial features from the list:

an estimate of direction of arrival of the near-field speech;

a calculation of a ratio of high pass-filtered energies of the at least two microphone signals;

an inverse signal-to-noise ratio of the near-field speech signal;

a zero-crossing rate of the at least two microphone signals;

a long-term average level of the background noise;

a multi-level discretized long-term background noise level; and

a multi-level discretized long-term signal-to-noise ratio of the near-field speech signal.

11. The method of claim 1 , further comprising:

calculating a ratio of high pass-filtered energies of the at least two microphone signals; and

detecting a presence of unvoiced speech using the calculation of the ratio of the high pass-filtered energies of the at least two microphone signals.

12. The method of claim 1 , further comprising:

using the estimated near-field speech presence probability to dynamically control an aggressiveness of said reducing the background noise by using the estimated power spectral density.

13. An audio device, comprising:

an array of at least two microphones;

a voice processing system, configured to:

obtain a multi-dimensional spatial feature vector comprising at least a correlation of the at least two microphones and a calculation of at least one ratio of energies of the at least two microphones;

use the multi-dimensional feature vector to estimate an energy of near-field speech and to estimate an energy of background noise;

use a ratio of the near-field speech energy estimate and the background noise energy estimate to estimate a probability of a presence of the near-field speech;

adaptively combine signals from the at least two microphones based on the estimated near-field speech presence probability to provide a combined output signal comprising a near-field speech signal and a residual background noise signal;

estimate a power spectral density of the residual background noise signal present at the combined output signal using the estimated near-field speech presence probability; and

reduce the background noise by using the estimated power spectral density.

14. The audio device of claim 13 , wherein to reduce the background noise, the voice processing system is configured to:

calculate a time-frequency dependent gain function based on the estimated residual background noise power spectral density and apply the calculated gain on the combined output signal to reduce the background noise.

15. The audio device of claim 13 , wherein the voice processing system is further configured to:

adaptively control the estimate of the power spectral density of the residual background noise signal when at least one of the following varies: the background noise energy and the near-field speech energy.

16. The audio device of claim 15 , wherein a level of the background noise is discretized as N-state noise level modes and M-state SNR modes, wherein N and M are integers greater than one.

17. The audio device of claim 16 , wherein the voice processing system is further configured to:

dynamically modify thresholds as a function of prevailing conditions of the background noise level and SNR; and

compare the spatial features of the multi-dimensional feature vector with the dynamically modified thresholds to generate control signals.

18. The audio device of claim 13 , wherein the voice processing system is further configured to:

determine that the background noise comprises wind noise;

select as an output signal one of the following based on diversity of energy of the wind noise among the at least two microphones:

the combined output signal; and

the signal from one of the at least two microphones;

estimate a power spectral density of the wind noise present at the selected output signal; and

reduce the wind noise by using the estimated power spectral density of the wind noise.

19. The audio device of claim 18 , wherein the power spectral density of the wind noise is estimated using the multi-dimensional spatial feature vector.

20. The audio device of claim 19 , wherein the power spectral density of the wind noise is estimated differently depending on whether acoustic noise or wind noise is present.

21. The audio device of claim 13 , wherein the array of at least two microphones are spaced at least 100 millimeters apart.

22. The audio device of claim 13 , wherein the multi-dimensional spatial feature vector further comprises one or more spatial features from the list:

an estimate of direction of arrival of the near-field speech;

a calculation of a ratio of high pass-filtered energies of the at least two microphone signals;

an inverse signal-to-noise ratio of the near-field speech signal;

a zero-crossing rate of the at least two microphone signals;

a long-term average level of the background noise;

a multi-level discretized long-term background noise level; and

a multi-level discretized long-term signal-to-noise ratio of the near-field speech signal.

23. The audio device of claim 13 , wherein the voice processing system is further configured to:

calculate a ratio of high pass-filtered energies of the at least two microphone signals; and

detect a presence of unvoiced speech using the calculation of the ratio of the high pass-filtered energies of the at least two microphone signals.

24. The audio device of claim 13 , wherein the voice processing system is further configured to:

use the estimated near-field speech presence probability to dynamically control an aggressiveness at which the background noise is reduced by using the estimated power spectral density.

25. A non-transitory computer-readable medium having instructions stored thereon that are capable of causing or configuring an audio device having an array of at least two microphones and a voice processing system to perform operations comprising:

obtaining a multi-dimensional spatial feature vector comprising at least a correlation of the at least two microphones and a calculation of at least one ratio of energies of the at least two microphones;

using the multi-dimensional feature vector to estimate an energy of near-field speech and to estimate an energy of background noise;

using a ratio of the near-field speech energy estimate and the background noise energy estimate to estimate a probability of a presence of the near-field speech;

adaptively combining signals from the at least two microphones based on the estimated near-field speech presence probability to provide a combined output signal comprising a near-field speech signal and a residual background noise signal;

estimating a power spectral density of the residual background noise signal present at the combined output signal using the estimated near-field speech presence probability; and

reducing the background noise by using the estimated power spectral density.

26. The non-transitory computer-readable medium of claim 25 , wherein said reducing the background noise comprises calculating a time-frequency dependent gain function based on the estimated residual background noise power spectral density and applying the calculated gain on the combined output signal to reduce the background noise.

27. The non-transitory computer-readable medium of claim 25 , wherein instructions stored thereon are capable of causing or configuring the audio device to perform further operations comprising:

adaptively controlling said estimating the power spectral density of the residual background noise signal when at least one of the following varies: the background noise energy and the near-field speech energy.

28. The non-transitory computer-readable medium of claim 25 , wherein instructions stored thereon are capable of causing or configuring the audio device to perform further operations comprising:

determining that the background noise comprises wind noise;

selecting as an output signal one of the following based on diversity of energy of the wind noise among the at least two microphones:

the combined output signal; and

the signal from one of the at least two microphones;

estimating a power spectral density of the wind noise present at the selected output signal; and

reducing the wind noise by using the estimated power spectral density of the wind noise.

29. The non-transitory computer-readable medium of claim 25 , wherein the multi-dimensional spatial feature vector further comprises one or more spatial features from the list:

an estimate of direction of arrival of the near-field speech;

a calculation of a ratio of high pass-filtered energies of the at least two microphone signals;

an inverse signal-to-noise ratio of the near-field speech signal;

a zero-crossing rate of the at least two microphone signals;

a long-term average level of the background noise;

a multi-level discretized long-term background noise level; and

a multi-level discretized long-term signal-to-noise ratio of the near-field speech signal.

30. The non-transitory computer-readable medium of claim 25 , wherein instructions stored thereon are capable of causing or configuring the audio device to perform further operations comprising:

calculating a ratio of high pass-filtered energies of the at least two microphone signals; and

detecting a presence of unvoiced speech using the calculation of the ratio of the high pass-filtered energies of the at least two microphone signals.

31. The non-transitory computer-readable medium of claim 25 , wherein instructions stored thereon are capable of causing or configuring the audio device to perform further operations comprising:

using the estimated near-field speech presence probability to dynamically control an aggressiveness of said reducing the background noise by using the estimated power spectral density.