Wind noise suppression system
View Patent ↗The invention includes a method, apparatus, and computer program to selectively suppress wind noise while preserving narrow-band signals in acoustic data. Sound from one or several microphones is digitized into binary data. A time-frequency transform is applied to the data to produce a series of spectra. The spectra are analyzed to detect the presence of wind noise and narrow band signals. Wind noise is selectively suppressed while preserving the narrow band signals. The narrow band signal is interpolated through the times and frequencies when it is masked by the wind noise. A time series is then synthesized from the signal spectral estimate that can be listened to. This invention overcomes prior art limitations that require more than one microphone and an independent measurement of wind speed. Its application results in good-quality speech from data severely degraded by wind noise.
1. A method for attenuating wind noise in a signal detected by a sound detector device, comprising:
performing a time-frequency transform on said signal to obtain transformed data;
performing signal analysis, by a signal analyzer implemented in hardware or program code embodied in a computer-readable storage medium, to identify signal peaks in said transformed data and determine that the signal peaks include a wind noise peak indicating that wind noise is present in the signal, where wind noise comprises noise caused by wind pressure fluctuations associated with wind striking a portion of the sound detector device, where performing signal analysis comprises:
identifying non-wind-noise peaks from among the signal peaks as sharper and narrower than a selected criteria; and
selecting the wind noise peak from among the signal peaks other than the non-wind-noise peaks; and
attenuating the wind noise peak identified in the signal analysis.
2. The method of claim 1 where performing signal analysis further comprises:
assigning evidence weights based on features of a spectrum of the transformed data; and
processing said evidence weights to determine wind noise presence.
3. The method of claim 2 where processing said evidence weights uses a fuzzy classifier.
4. The method of claim 2 where processing said evidence weights uses an artificial neural network.
5. The method of claim 1 where performing signal analysis further comprises:
identifying a non-wind-noise peak in said signal peaks when a Signal to Noise Ratio (SNR) exceeds a peak threshold.
6. The method of claim 1 where identifying comprises measuring peak widths by taking an average difference between a highest point and its neighboring points on each side.
7. The method of claim 1 where identifying further comprises:
identifying a data point in the transformed data as one of said signal peaks if it is greater in value than both of its neighboring data points; and
classifying said data point as a non-wind-noise peak if it is greater in value than the value of two data points, in either direction a number of units away, by a decibel threshold.
8. The method of claim 7 wherein said number of units is two.
9. The method of claim 7 wherein said decibel threshold is 7 dB.
10. The method of claim 1 where performing signal analysis further comprises:
determining whether there is a harmonic relationship between selected signal peaks.
11. The method of claim 10 where determining whether there is a harmonic relationship further comprises:
applying a direct cosine transform (DCT) to said transformed data along a frequency axis to produce a normalized DCT, wherein said normalized DCT is normalized by the first value of the DCT transform; and
determining whether there is a maximum at a value in said normalized DCT at a value of a pitch period corresponding to a non-wind-noise peak.
12. The method of claim 1 , further comprising:
determining stability of the signal peaks by comparing the signal peaks in said transformed data to signal peaks from prior transformed data; and
identifying stable peaks as non-wind-noise peaks.
13. The method of claim 1 , further comprising:
determining differences in phase and amplitude of the signal peaks from signals from a plurality of microphones; and
identifying a wind noise peak as having phase and amplitude differences exceeding a difference threshold.
14. The method of claim 1 where performing signal analysis further comprises:
measuring a rate of variation of a lower frequency portion of said transformed data.
15. The method of claim 1 where performing the time-frequency transform further comprises:
performing condition operations on said signal.
16. The method of claim 15 wherein said condition operations comprise:
pre-filtering.
17. The method of claim 15 wherein said condition operations comprise:
shading.
18. The method of claim 1 where performing the time-frequency transform uses a short-time Fourier transform.
19. The method of claim 1 where performing the time-frequency transform uses a bank of filter analysis.
20. The method of claim 1 where performing the time-frequency transform uses a discrete wavelet transform.
21. The method of claim 1 where attenuating wind noise further comprises:
suppressing portions of the transformed data dominated by wind noise; and
preserving portions of the transformed data dominated by the non-wind-noise peaks.
22. The method of claim 21 further comprising:
generating a low-noise version of the transformed data.
23. The method of claim 1 , further comprising:
constructing a wind noise attenuated time series using an inverse Fourier transform.
24. The method of claim 1 , further comprising:
sampling said signal to obtain sampled data; and
creating data buffers from said sampled data.
25. The method of claim 24 where performing a time-frequency transform performs transformation on each of said buffers as it is created.
26. The method of claim 1 , further comprising:
performing reconstruction of the signal by interpolation or extrapolation through a time or frequency region masked by wind noise.
27. The method of claim 1 , further comprising:
detecting a transient signal in said transformed data.
28. The method of claim 27 , further comprising:
averaging acoustic power in a sliding window for frequency bands in said transformed data; and
declaring presence of a transient signal when average acoustic power over a pre-determined number of the frequency bands exceeds background noise by more than a decibel (dB) threshold.
29. The method of claim 28 where said decibel threshold is between 6 to 12 dB.
30. The method of claim 1 , further comprising:
performing curve fitting to a lower frequency portion of said transformed data; and
comparing curve parameters to a plurality of pre-defined thresholds associated with wind noise to determine whether the signal includes wind noise.
31. The method of claim 30 wherein said curve fitting is performed by fitting a straight line to the lower frequency portion.
32. The method of claim 30 wherein said curve parameters comprise:
a slope value; and
an intersection point;
where comparing curve parameters further comprises identifying wind noise in the signal based on a comparison between the slope value and a slope value threshold associated with wind noise, and a comparison between the intersection point and an intersection point threshold associated with wind noise.
33. The method of claim 1 , further comprising:
obtaining the signal from a single microphone source.
34. The method of claim 1 , further comprising:
estimating background noise in the signal;
suspending estimating the background noise upon detection of a transient signal; and
resuming estimating the background noise once said transient signal passes.
35. The method of claim 1 , further comprising:
estimating background noise in the signal; and
dynamically tracking said background noise.
36. The method of claim 1 , further comprising;
estimating background noise in the signal during boot-up of a computer; and
using the background noise to alternate the wind noise.
37. The method of claim 1 , further comprising:
estimating background noise in the signal; and
using the background noise to alternate the wind noise.
38. The method of claim 1 , further comprising:
constructing a wind noise attenuated time series using the transformed data.
39. An apparatus for suppressing wind noise, comprising:
a time-frequency transform component configured to transform a time-based signal to frequency-based data;
a signal analyzer configured to identify signal peaks in said frequency-based data and determine that said signal peaks include a wind noise peak indicating that wind noise is present, where wind noise comprises noise caused by wind pressure fluctuations associated with wind striking a portion of a sound detector device that detected the time-based signal, where the signal analyzer comprises hardware or a computer-readable storage medium that stores program code executable by a processor, where the signal analyzer is further configured to:
identify non-wind-noise peaks when there is a harmonic relationship between the signal peaks; and
select the wind noise peak from among the signal peaks other than the non-wind-noise peaks; and
a wind noise attenuation component configured to attenuate the wind noise peak identified by the signal analyzer.
40. The apparatus of claim 39 wherein said signal analyzer is configured to:
assign evidence weights based on features of a spectrum of the frequency-based data; and
process said evidence weights to determine wind noise presence.
41. The apparatus of claim 40 wherein said signal analyzer is configured to use a fuzzy classifier to process said evidence weights.
42. The apparatus of claim 40 wherein said signal analyzer is configured to use an artificial neural network to process said evidence weights.
43. The apparatus of claim 39 wherein said signal analyzer is further configured to identify non-wind-noise peaks in the signal peaks as sharper and narrower than a certain criteria.
44. The apparatus of claim 43 wherein said signal analyzer is configured to measure peak widths by taking an average difference between a highest point and its neighboring points on each side.
45. The apparatus of claim 43 wherein said signal analyzer is configured to:
identify a data point in the frequency-based data as one of the signal peaks if it is greater in value than both of its neighboring data points; and
classify said data point as a non-wind-noise peak if it is greater in value than the value of two data points, in either direction a number of units away, by a decibel threshold.
46. The apparatus of claim 45 wherein said number of units is two.
47. The apparatus of claim 45 wherein said decibel threshold is 7 dB.
48. The apparatus of claim 39 wherein said signal analyzer is further configured to identify a non-wind-noise peak in the signal peaks that has a Signal to Noise Ratio (SNR) exceeding a peak threshold.
49. The apparatus of claim 39 wherein said signal analyzer is configured to determine when there is a harmonic relationship by:
applying a direct cosine transform (DCT) to said frequency-based data along the frequency axis to produce a normalized DCT, wherein said normalized DCT is normalized by the first value of the DCT transform; and
determining whether there is a maximum at the value in said normalized DCT at a value of a pitch period corresponding to a non-wind-noise peak.
50. The apparatus of claim 39 wherein said signal analyzer is configured to analyze by:
determining stability of the signal peaks by comparing the signal peaks in said frequency-based data to signal peaks from prior frequency-based data; and
identifying stable peaks as non-wind-noise peaks.
51. The apparatus of claim 39 wherein said signal analyzer is configured to analyze by:
determining differences in phase and amplitude of the signal peaks from signals from a plurality of microphones; and
identifying a wind noise peak as having phase and amplitude differences exceeding a difference threshold.
52. The apparatus of claim 39 wherein said signal analyzer is configured to analyze by:
measuring a rate of variation of a lower frequency portion of said frequency based data.
53. The apparatus of claim 39 wherein said time-frequency transform component is configured to perform condition operations on said signal.
54. The apparatus of claim 53 wherein said condition operations comprise:
pre-filtering.
55. The apparatus of claim 53 wherein said condition operations comprise:
shading.
56. The apparatus of claim 39 wherein said time-frequency transform component is configured to use a short-time Fourier transform.
57. The apparatus of claim 39 wherein said time-frequency transform component is configured to use a bank of filter analysis.
58. The apparatus of claim 39 wherein said time-frequency transform component is configured to use a discrete wavelet transform.
59. The apparatus of claim 39 wherein said wind noise attenuation component is configured to attenuate wind noise by:
suppressing portions of the frequency-based data dominated by wind noise; and
preserving portions of the frequency-based data dominated by the non-wind-noise peaks.
60. The apparatus of claim 59 , where said wind noise attenuation component is configured to attenuate wind noise by generating a low-noise version of the frequency-based data.
61. The apparatus of claim 39 , further comprising a time series synthesis component configured to construct a wind noise attenuated time series using an inverse Fourier transform.
62. The apparatus of claim 39 , further comprising:
a sampling component configured to sample said signal to obtain sampled data and create data buffers from said sampled data.
63. The apparatus of claim 62 wherein said time-frequency transform performs transformation on each of said buffers as it is created.
64. The apparatus of claim 39 , further comprising:
a reconstruction component configured to reconstruct the signal by interpolation or extrapolation through the time or frequency regions that were masked by wind noise.
65. The apparatus of claim 39 , further comprising:
a detecting component configured to detect a transient signal in said frequency-based data.
66. The apparatus of claim 65 wherein said detecting component is configured to detect by:
averaging acoustic power in a sliding window for frequency bands in said frequency-based data; and
declaring presence of a transient signal when average acoustic power within a pre-determined number of frequency bands exceed the background noise by more than a decibel (dB) threshold.
67. The apparatus of claim 66 wherein said decibel threshold is between 6 to 12 dB.
68. The apparatus of claim 39 wherein said signal analyzer is configured to:
perform curve fitting to a lower frequency portion of said frequency-based data; and
compare curve parameters to a plurality of pre-defined thresholds associated with wind noise to determine whether the signal includes wind noise.
69. The apparatus of claim 68 wherein said curve fitting is performed by fitting a straight line to the lower frequency portion.
70. The apparatus of claim 68 wherein said curve parameters comprise:
a slope value; and
an intersection point;
where said signal analyzer is configured to identify wind noise in the signal based on a comparison between the slope value and a slope value threshold associated with wind noise, and a comparison between the intersection point and an intersection point threshold associated with wind noise.
71. The apparatus of claim 39 wherein said signal is from a single microphone source.
72. The apparatus of claim 39 , further comprising:
an estimating component configured to estimate background noise in the signal wherein the background noise is used to attenuate wind noise.
73. The apparatus of claim 39 , further comprising:
a time-series synthesis component configured to construct a wind noise attenuated time series using the frequency based data.
74. A computer program product comprising:
a computer usable storage medium having computer readable program code embodied therein configured for suppressing wind noise, comprising:
computer readable code configured to cause a computer to perform a time-frequency transform on a signal to obtain transformed data;
computer readable code configured to cause the computer to perform signal analysis to identify signal peaks in said transformed data and determine that the signal peaks include a wind noise peak indicating that wind noise is present in the signal, where wind noise comprises noise caused by wind pressure fluctuations associated with wind striking a portion of a sound detector device that detected the signal;
computer readable code configured to cause the computer to determine differences in phase and amplitude of the signal peaks from signals from multiple microphones;
computer readable code configured to cause the computer to identify wind noise peaks when the differences exceed a difference threshold; and
computer readable code configured to cause the computer to attenuate the wind noise.
75. The computer program product of claim 74 where said computer readable program code further comprises:
computer readable code configured to cause the computer to assign evidence weights for identifying wind noise; and
computer readable code configured to cause the computer to process said evidence weights to determine that wind noise is present.
76. The computer program product of claim 75 wherein said computer readable program code uses a fuzzy classifier.
77. The computer program product of claim 75 wherein said computer readable program code uses an artificial neural network.
78. The computer program product of claim 74 where said computer readable program code further comprises:
computer readable code configured to cause the computer to identify non-wind-noise peaks in the signal peaks as sharper and narrower than a certain criteria.
79. The computer program product of claim 78 wherein said computer readable program code is configured to measure peak widths by taking an average difference between a highest point and its neighboring points on each side.
80. The computer program product of claim 78 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to identify a data point in the transformed data as one of the signal peaks if it is greater in value than both of its neighboring data points; and
computer readable code configured to cause the computer to classify said data point as a non-wind-noise peak if it is greater in value than the value of two data points, in either direction a number of units away, by a decibel threshold.
81. The computer program product of claim 80 wherein said number of units is two.
82. The computer program product of claim 80 wherein said decibel threshold is 7 dB.
83. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to determine whether there is a harmonic relationship between selected signal peaks.
84. The computer program product of claim 83 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to apply a direct cosine transform (DCT) to said transformed data along a frequency axis to produce a normalized DCT, wherein said normalized DCT is normalized by the first value of the DCT transform; and
computer readable code configured to cause the computer to determine whether there is a maximum at a value in said normalized DCT at a value of a pitch period corresponding to a non-wind-noise peak.
85. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to determine stability of the signal peaks by comparing the signal peaks in said transformed data to signal peaks from prior transformed data; and
computer readable code configured to cause the computer to identify stable peaks as non-wind-noise peaks.
86. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to analyze features of said transformed data by identifying a non-wind-noise peak in the signal peaks when a Signal to Noise Ratio (SNR) exceeds a peak threshold.
87. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to measure a rate of variation of a lower frequency portion of said transformed data.
88. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to perform condition operations on said signal.
89. The computer program product of claim 88 wherein said condition operations comprise:
pre-filtering.
90. The computer program product of claim 88 wherein said condition operations comprise:
shading.
91. The computer program product of claim 74 wherein said computer readable program code uses a short-time Fourier transform.
92. The computer program product of claim 74 wherein said computer readable program code uses a bank of filter analysis.
93. The computer program product of claim 74 wherein said computer readable program code uses a discrete wavelet transform.
94. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to suppress portions of the transformed data dominated by wind noise; and
computer readable code configured to cause the computer to preserve portions of the transformed data dominated by non-wind-noise peaks.
95. The computer program product of claim 94 , further comprising:
computer readable code configured to cause the computer to generate a low-noise version of the transformed data.
96. The computer program product of claim 74 wherein said computer readable program code is configured to construct a wind noise attenuated time series using an inverse Fourier transform.
97. The computer program product of claim 74 , further comprising:
computer readable code configured to cause the computer to sample said signal to obtain sampled data; and
computer readable code configured to cause the computer to create data buffers from said sampled data.
98. The computer program product of claim 97 wherein said computer readable code configured to cause the computer to perform time-frequency transform causes the computer to perform transformation on each of said buffers as it is created.
99. The computer program product of claim 74 , further comprising:
computer readable code configured to cause the computer to perform reconstruction of the signal by interpolation or extrapolation through a time or frequency region masked by wind noise.
100. The computer program product of claim 74 , further comprising:
computer readable code configured to cause the computer to estimate background noise in said transformed data, wherein said background noise is used to attenuate wind noise.
101. The computer program product of claim 100 further comprising:
computer readable code configured to cause the computer to detect a transient signal in said transformed data.
102. The computer program product of claim 101 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to average acoustic power in a sliding window for frequency bands in said transformed data; and
computer readable code configured to cause the computer to declare presence of a transient signal when average acoustic power over a pre-determined number of frequency bands exceeds the background noise by more than a decibel (dB) threshold decibel.
103. The computer program product of claim 102 wherein said threshold is between 6 to 12 dB.
104. The computer program product of claim 74 wherein said computer readable program code further comprises:
computer readable code configured to cause the computer to perform curve fitting to a lower frequency portion of said transformed data; and
computer readable code configured to cause the computer to compare curve parameters to a plurality of pre-defined thresholds associated with wind noise to determine whether the signal includes wind noise.
105. The computer program product of claim 104 wherein said curve fitting is performed by fitting a straight line to the lower frequency portion.
106. The computer program product of claim 104 wherein said curve parameters comprise:
a slope value; and
an intersection point;
wherein said computer readable program code further comprises computer readable code configured to cause the computer to identify wind noise in the signal based on a comparison between the slope value and a slope value threshold associated with wind noise, and a comparison between the intersection point and an intersection point threshold associated with wind noise.
107. The computer program product of claim 74 wherein the signal is from a single microphone source.
108. The computer program product of claim 74 , further comprising:
computer readable code configured to construct a wind noise attenuated time series from the transformed data.
109. An apparatus for suppressing wind noise, comprising:
a time-frequency transform component configured to transform a time-based signal to frequency-based data;
a signal analyzer configured to identify signal peaks in said frequency-based data and determine that said signal peaks include a wind noise peak indicating that wind noise is present, where wind noise comprises noise caused by wind pressure fluctuations associated with wind striking a portion of a sound detector device that detected the time-based signal, where the signal analyzer comprises hardware or a computer-readable storage medium that stores program code executable by a processor, where the signal analyzer is further configured to:
identify non-wind-noise peaks from among the signal peaks as sharper and narrower than a selected criteria; and
select the wind noise peak from among the signal peaks other than the non-wind-noise peaks;
a wind noise attenuation component configured to attenuate the wind noise peak identified by the signal analyzer.
110. The apparatus of claim 109 wherein said signal analyzer is further configured to measure peak widths of the signal peaks by taking an average difference between a highest point and its neighboring points on each side.
111. The apparatus of claim 109 wherein said signal analyzer is configured to:
identify a data point in the frequency-based data as one of the signal peaks if it is greater in value than both of its neighboring data points; and
classify said data point as a non-wind-noise peak if it is greater in value than the value of two data points, in either direction a number of units away, by a decibel threshold.
112. The apparatus of claim 111 wherein said number of units is two.
113. The apparatus of claim 111 wherein said decibel threshold is 7 dB.
114. A method for attenuating wind noise in a signal, comprising:
fitting a line to a portion of a frequency spectrum of the signal;
calculating a slope of the line;
identifying, by a signal analyzer implemented in hardware or program code embodied in a computer usable storage device, whether the portion of the signal contains wind noise based on a comparison between the slope of the line and a slope threshold associated with wind noise, where wind noise comprises noise caused by wind pressure fluctuations associated with wind striking a portion of a sound detector device that detected the signal; and
attenuating wind noise in the portion of the signal when wind noise is identified by the comparison.
115. The method of claim 114 , where identifying comprises identifying whether the portion of the signal contains wind noise based on the comparison between the slope of the line and the slope threshold associated with wind noise, and a comparison between an intersection point of the line and an intersection point threshold associated with wind noise.