Hybrid noise suppression
A method for hybrid noise suppression. The method involves obtaining a noisy audio input signal, generating a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression, and outputting the noise-suppressed audio output signal.
1. A method for hybrid noise suppression, comprising:
obtaining a noisy audio input signal;
generating a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression, wherein performing further comprises:
obtaining a processed audio signal by scaling the noisy audio input signal in a first frequency domain using the noise spectrum estimate-based noise suppression;
scaling the processed audio signal in a second frequency domain using the machine learning-based noise suppression; and
outputting the noise-suppressed audio output signal.
2. The method of claim 1 , wherein obtaining the processed audio signal by scaling the noisy audio input signal in the first frequency domain using the noise spectrum estimate-based noise suppression comprises:
transforming the noisy audio input signal from a time domain into the first frequency domain comprising a plurality of frequency sub-bands;
obtaining a noise estimate in the plurality of frequency sub-bands;
obtaining the processed audio signal by removing the noise estimate from the noisy audio input signal for the plurality of frequency sub-bands; and
transforming, after obtaining the processed audio signal, the processed audio signal from the first frequency domain to the time domain.
3. The method of claim 1 , wherein the obtaining the noise spectrum estimate comprises:
identifying an interval of minimum energy in the noisy audio input signal.
4. The method of claim 1 , wherein obtaining the noise-suppressed audio output signal by scaling the processed audio signal in the second frequency domain using the machine learning-based noise suppression comprises:
transforming the noisy audio input signal from a time domain into the second frequency domain comprising a plurality of frequency sub-bands;
generating a feature vector from the noisy audio input signal in the plurality of frequency sub-bands;
obtaining a set of sub-band gain values by applying a neural network model to the feature vector;
transforming the processed audio signal from the time domain into the second frequency domain comprising the plurality of frequency sub-bands;
scaling, according to the sub-band gain values, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio output signal; and
transforming, after scaling according to the sub-band gain values, the noise-suppressed audio output signal from the second frequency domain to the time domain.
5. The method of claim 4 , further comprising, prior to transforming the noisy audio input signal from the time domain to the second frequency domain:
performing an echo cancellation on the noisy audio input signal in the first frequency domain; and
transforming, after the echo cancellation, the noisy audio input signal from the first frequency domain to the time domain.
6. The method of claim 4 , wherein the neural network model is a deep neural network.
7. The method of claim 6 , wherein the deep neural network comprises gated recurrent units.
8. The method of claim 1 , wherein obtaining the noise-suppressed audio output signal by scaling the processed audio signal in the second frequency domain using the machine learning-based noise suppression comprises:
transforming the processed audio signal from the time domain into the second frequency domain comprising a plurality of frequency sub-bands;
generating a feature vector from the processed audio signal in the plurality of frequency sub-bands;
obtaining a set of sub-band gain values by applying a neural network model to the feature vector;
scaling, according to the sub-band gain values, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio output signal; and
transforming, after scaling according to the sub-band gain values, the noise-suppressed audio output signal from the second frequency domain to the time domain.
9. A method for hybrid noise suppression, comprising:
obtaining a noisy audio input signal;
generating a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression; and
outputting the noise-suppressed audio output signal,
wherein performing the combined noise spectrum estimate-based and machine learning-base noise suppression comprises:
transforming the noisy audio input signal from a time domain into the frequency domain comprising a plurality of frequency sub-bands;
obtaining a noise estimate in the plurality of frequency sub-bands;
determining a first set of sub-band gain values configured to remove the noise estimate from the noisy audio input signal;
generating a feature vector from the noisy audio input signal in the plurality of frequency sub-bands;
determining a second set of sub-band gain values by applying a neural network model to the feature vector;
combining the first set of sub-band gain values and the second set of sub-band gain values to form a combined set of sub-band gain values;
scaling, according to the combined set of sub-band gain values, the noisy audio input signal in the plurality of frequency sub-bands to obtain the noise-suppressed audio output signal; and
transforming, after scaling, the noise-suppressed audio output signal from the frequency domain to the time domain.
10. A system for hybrid noise suppression, comprising:
an audio source; and
a computing system configured to:
obtain a noisy audio input signal from the audio source;
generate a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression, wherein performing the combination further comprises:
obtaining a processed audio signal by scaling the noisy audio input signal in a first frequency domain using the noise spectrum estimate-based noise suppression, and
obtaining the noise-suppressed audio output signal by scaling the processed audio signal in a second frequency domain using the machine learning-based noise suppression; and
output the noise-suppressed audio output signal.
11. The system of claim 10 ,
wherein the system is a teleconferencing device, and
wherein the audio source is a microphone.
12. The system of claim 10 , wherein obtaining the processed audio signal by scaling the noisy audio input signal in the first frequency domain using the noise spectrum estimate-based noise suppression comprises:
transforming the noisy audio input signal from a time domain into the first frequency domain comprising a plurality of frequency sub-bands;
obtaining a noise estimate in the plurality of frequency sub-bands;
obtaining the processed audio signal by removing the noise estimate from the noisy audio input signal for the plurality of frequency sub-bands; and
transforming, after obtaining the processed audio signal, the processed audio signal from the first frequency domain to the time domain.
13. The system of claim 10 , wherein obtaining the noise-suppressed audio output signal by scaling the processed audio signal in the second frequency domain using the machine learning-based noise suppression comprises:
transforming the noisy audio input signal from a time domain into the second frequency domain comprising a plurality of frequency sub-bands;
generating a feature vector from the noisy audio input signal in the plurality of frequency sub-bands;
obtaining a set of sub-band gain values by applying a neural network model to the feature vector;
transforming the processed audio signal from the time domain into the second frequency domain comprising the plurality of frequency sub-bands;
scaling, according to the sub-band gain values, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio output signal; and
transforming, after scaling according to the sub-band gain values, the noise-suppressed audio output signal from the second frequency domain to the time domain.
14. The system of claim 13 , wherein the computing system is further configured to, prior to transforming the noisy audio input signal from the time domain to the second frequency domain:
perform an echo cancellation on the noisy audio input signal in the first frequency domain; and
transform, after the echo cancellation, the noisy audio input signal from the first frequency domain to the time domain.
15. The system of claim 13 , wherein the neural network model is a deep neural network.
16. The system of claim 10 , wherein obtaining the noise-suppressed audio output signal by scaling the processed audio signal in the second frequency domain using the machine learning-based noise suppression comprises:
transforming the processed audio signal from the time domain into the second frequency domain comprising a plurality of frequency sub-bands;
generating a feature vector from the processed audio signal in the plurality of frequency sub-bands;
obtaining a set of sub-band gain values by applying a neural network model to the feature vector;
scaling, according to the sub-band gain values, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio output signal; and
transforming, after scaling according to the sub-band gain values, the noise-suppressed audio output signal from the second frequency domain to the time domain.
17. A system for hybrid noise suppression, comprising:
an audio source; and
a computing system configured to:
obtain a noisy audio input signal from the audio source;
generate a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression; and
output the noise-suppressed audio output signal,
wherein performing the combined noise spectrum estimate-based and machine learning-base noise suppression comprises:
transforming the noisy audio input signal from a time domain into the frequency domain comprising a plurality of frequency sub-bands;
obtaining a noise estimate in the plurality of frequency sub-bands;
determining a first set of sub-band gain values configured to remove the noise estimate from the noisy audio input signal;
generating a feature vector from the noisy audio input signal in the plurality of frequency sub-bands;
determining a second set of sub-band gain values by applying a neural network model to the feature vector;
combining the first set of sub-band gain values and the second set of sub-band gain values to form a combined set of sub-band gain values;
scaling, according to the combined set of sub-band gain values, the noisy audio input signal in the plurality of frequency sub-bands to obtain the noise-suppressed audio output signal; and
transforming, after scaling, the noise-suppressed audio output signal from the frequency domain to the time domain.
18. A non-transitory computer readable medium comprising computer readable program code for hybrid noise suppression, the computer readable program code causing a computer system to:
obtain a noisy audio input signal;
generate a noise-suppressed audio output signal by performing a combination of a noise spectrum estimate-based noise suppression and a machine learning-based noise suppression, wherein performing further comprises:
obtaining a processed audio signal by scaling the noisy audio input signal in a first frequency domain using the noise spectrum estimate-based noise suppression;
scaling the processed audio signal in a second frequency domain using the machine learning-based noise suppression; and
output the noise-suppressed audio output signal.