Granted Patent B2

US 12,456,476 · App. 18/081,492 · Granted Oct 28, 2025

Inventors: Chandan Karadagur Ananda Reddy (Cupertino, CA); Navin Chatlani (Palo Alto, CA)

Assignee: Google LLC

G10L21/0232G10L21/0224G10L25/60G10L2021/02163

Abstract

Claims (72)

1. A computer-implemented method comprising:

receiving, by one or more processors, a current time frame of speech data in a time domain;

transforming the current time frame to a current frequency frame of the speech data in a frequency domain;

determining, using a noise classifier implemented by the one or more processors, a classification of noise content in the current frequency frame based at least on noise content determined in the current frequency frame, wherein the classification is selected from at least two classifications detectable by the noise classifier, wherein the two classifications include a first classification associated with creating and applying a current noise suppression mask to the current frequency frame, and a second classification associated with applying a previously-created noise suppression mask to the current frequency frame without creating the current noise suppression mask;

in response to determining the first classification:

creating, by the one or more processors, the current noise suppression mask for the current frequency frame based on the noise content in the current frequency frame, wherein creating the current noise suppression mask includes determining one or more gain functions associated with one or more frequency bands of the current frequency frame; and

applying, by the one or more processors, the current noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain a current noise-suppressed frequency frame of the speech data;

in response to determining the second classification:

selecting, by the one or more processors, the previously-created noise suppression mask, wherein the previously-created noise suppression mask was created and applied based on a prior frequency frame of the speech data; and

applying, by the one or more processors, the previously-created noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain the current noise-suppressed frequency frame of the speech data, without creating the current noise suppression mask;

transforming the current noise-suppressed frequency frame of the speech data to a current noise-suppressed time frame of the speech data that is in the time domain; and

outputting the current noise-suppressed time frame of the speech data that provides audio output having reduced noise.

2. The method of claim 1 , wherein the noise classifier uses a machine-learning model that is trained to determine the classification of noise content in frequency frames provided as input to the machine-learning model.

3. The method of claim 2 , wherein the machine-learning model is trained based on speech data that does not include noise and noise data that includes background noise.

4. The method of claim 2 , wherein the machine-learning model is trained using a speech quality predictor that estimates a quality of the speech data.

5. The method of claim 1 , wherein determining the classification of the noise content in the current frequency frame is based on a determined magnitude of the noise content determined in the current frequency frame.

6. The method of claim 1 , wherein determining the classification of the noise content in the current frequency frame is based on a determined rate of change of the noise content in the current time frame relative to determined noise content in one or more previous time frames.

7. The method of claim 1 , further comprising:

receiving a second time frame of the speech data in the time domain;

transforming the second time frame to a second frequency frame of the speech data in the frequency domain;

determining, using the noise classifier, whether to create a second noise suppression mask for the second frequency frame based at least on noise content determined in the second frequency frame;

in response to determining to create the second noise suppression mask:

creating a second noise suppression mask based on the second frequency frame, wherein creating the second noise suppression mask includes determining one or more gain functions associated with one or more frequency bands of the second frequency frame; and

applying the second noise suppression mask to the second frequency frame to suppress the noise content in the second frequency frame and obtain a second noise-suppressed frequency frame;

in response to determining to not create the second noise suppression mask:

applying the current noise suppression mask to the second frequency frame to suppress the noise content in the second frequency frame and obtain the second noise-suppressed frequency frame of the speech data, without the creating and the applying of the second noise suppression mask;

transforming the second noise-suppressed frequency frame to a second noise-suppressed time frame; and

outputting the second noise-suppressed time frame of the speech data.

8. The method of claim 1 , wherein determining the classification of the noise content in the current frequency frame is based on the determined noise content in the current frequency frame and determined noise content in one or more previous frequency frames.

9. The method of claim 1 , wherein applying the current noise suppression mask to the current frequency frame is performed without applying the previously-created noise suppression mask.

10. The method of claim 1 , wherein selecting the previously-created noise suppression mask includes selecting a most recent previously-created noise suppression mask created for a most recent prior frequency frame of the speech data.

11. The method of claim 1 , wherein the noise classifier includes a statistical classifier, wherein the statistical classifier:

compares data in the current frequency frame to previous data in one or more previous frequency frames using a similarity function or a distance function; and

based on the comparison, determines whether a rate of change of noise in the one or more previous frequency frames and the current frequency frame meets a threshold that indicates the classification of the noise content in the current frequency frame.

12. A computing device comprising:

a processor; and

a memory coupled to the processor, with instructions stored thereon that, when executed by the processor, cause the processor to perform operations comprising:

receiving, by the processor, a current time frame of speech data in a time domain after receiving a first time frame of the speech data in the time domain, wherein the current time frame and the first time frame include speech content, and wherein a first noise suppression mask is associated with the first time frame and created to suppress noise in the first time frame of the speech data;

transforming the current time frame to a current frequency frame of the speech data in a frequency domain;

determining, using a noise classifier, a classification of noise content in the current frequency frame based at least on noise content determined in the current frequency frame, wherein the classification is selected from at least two classifications detectable by the noise classifier, wherein the two classifications include a first classification associated with creating and applying a current noise suppression mask to the current frequency frame, and a second classification associated with applying a previously-created noise suppression mask to the current frequency frame without creating the current noise suppression mask;

in response to determining the first classification:

creating the current noise suppression mask for the current frequency frame based on the noise content in the current frequency frame, wherein creating the current noise suppression mask includes determining one or more gain functions associated with one or more frequency bands of the current frequency frame; and

applying the current noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain a current noise-suppressed frequency frame of the speech data;

in response to determining the second classification:

selecting the first noise suppression mask; and

applying the first noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain the current noise-suppressed frequency frame of the speech data, without the creating and the applying of the current noise suppression mask;

transforming the current noise-suppressed frequency frame of the speech data to a current noise-suppressed time frame of the speech data that is in the time domain; and

outputting the current noise-suppressed time frame of the speech data.

13. The computing device of claim 12 , wherein the noise classifier uses a machine-learning model that is trained to determine the classification of noise content in frequency frames provided as input to the machine-learning model.

14. The computing device of claim 12 , wherein the noise classifier includes a statistical classifier, wherein the statistical classifier:

compares data in the current frequency frame to previous data in one or more previous frames using a similarity function or a distance function; and

based on the comparison, determines whether a rate of change of noise in the current frequency frame and the one or more previous frames meets a threshold that indicates the classification of the noise content in the current frequency frame.

15. A device comprising:

at least one battery;

at least one microphone;

a communication circuit coupled to the battery and the microphone;

a processor coupled to the communication circuit, the battery and the microphone; and

a memory coupled to the processor, with instructions stored thereon that, when executed by the processor, cause the processor to perform operations comprising:

receiving, by the processor via the at least one microphone, a current time frame of speech data in a time domain;

transforming the current time frame of speech data to a current frequency frame of the speech data in a frequency domain;

determining, using a noise classifier, a classification of noise content in the current frequency frame based at least on noise content determined in the current frequency frame wherein the classification is selected from at least two classifications detectable by the noise classifier, wherein the two classifications include a first classification associated with creating and applying a current noise suppression mask to the current frequency frame, and a second classification associated with applying a previously-created noise suppression mask to the current frequency frame without creating the current noise suppression mask;

in response to determining the first classification:

creating the current noise suppression mask for the current frequency frame based on the noise content in the current frequency frame, wherein creating the current noise suppression mask includes determining one or more gain functions associated with one or more frequency bands of the current frequency frame; and

applying the current noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain a current noise-suppressed frequency frame of the speech data;

in response to determining the second classification:

selecting the previously-created noise suppression mask, wherein the previously-created noise suppression mask was created based on a prior frequency frame of the speech data; and

applying the previously-created noise suppression mask to the current frequency frame to suppress the noise content in the current frequency frame and obtain the current noise-suppressed frequency frame of the speech data, without creating the current noise suppression mask; and

transforming the current noise-suppressed frequency frame of the speech data to a current noise-suppressed time frame of the speech data that is in the time domain.

16. The device of claim 15 , wherein the noise classifier uses a machine-learning model that is trained to determine the classification of noise content in frequency frames provided as input to the machine-learning model.

17. The device of claim 15 , wherein the noise classifier includes a statistical classifier, wherein the statistical classifier:

compares data in the current frequency frame to previous data in one or more previous frequency frames using a similarity function or a distance function; and

based on the comparison, determines whether a rate of change of noise in the current frequency frame and the one or more previous frequency frames meets a threshold that indicates the classification of the noise content in the current frequency frame.

Assignments (1)

Continuity (1)

References Cited (26)