Federated learning for audio processing
A system performs federated learning and retraining of a machine learning model used for processing audio detected by a user device. The system uses both gradient data (which may correspond to false-rejects) and audio data (which may correspond to false-positives) received from devices. The system may also use a teacher model to produce labels for data in an automated fashion, thus allowing retraining to happen in an unsupervised manner.
1. A computer-implemented method, comprising:
receiving, from a first device, gradient data corresponding to operation, by the first device, of a first machine learning model configured to detect a representation of a wakeword;
receiving, from a second device, first audio data corresponding to detection of the wakeword by the second device, wherein the detection resulted from operation of the first machine learning model by the second device;
processing the first audio data using a second machine learning model configured to detect a representation of the wakeword;
determining, as a result of processing the first audio data using the second machine learning model, that the detection of the wakeword was a false-positive;
processing the first audio data and the gradient data to determine an updated first machine learning model; and
sending the updated first machine learning model to the first device and the second device.
2. The computer-implemented method of claim 1 , further comprising, by the first device:
detecting first audio using at least one microphone;
determining first input audio data representing the first audio;
processing the first input audio data using the first machine learning model to determine first model output data indicating an absence of the wakeword in the first audio;
detecting second audio using the at least one microphone;
determining second input audio data representing the second audio;
processing the second input audio data using the first machine learning model to determine second model output data indicating detection of the wakeword in the second audio;
determining the second audio was detected within a time threshold of the first audio;
based at least in part on the second audio being detected within the time threshold of the first audio, determining the gradient data;
determining metadata corresponding to the gradient data; and
sending the metadata.
3. The computer-implemented method of claim 1 , further comprising:
receiving, from the second device, first metadata corresponding to the first audio data;
performing automatic speech recognition (ASR) processing using the first audio data to determine ASR data;
performing natural language understanding (NLU) processing using the ASR data to determine NLU data; and
processing the first metadata, the ASR data, the NLU data, and the first audio data using a third machine learning model to determine label data corresponding to representation of the wakeword in the first audio data,
wherein determination of the updated first machine learning model further comprises processing the label data.
4. The computer-implemented method of claim 1 , further comprising:
receiving, from a third device, second gradient data;
determining the gradient data is associated with a first user characteristic;
determining the first audio data is associated with the first user characteristic; and
determining the second gradient data is associated with a second user characteristic different from the first user characteristic,
wherein determination of the updated first machine learning model is performed without involving the second gradient data in response to the second gradient data being associated with the second user characteristic.
5. A computer-implemented method, comprising:
receiving first parameter data corresponding to adjustment of at least one parameter of a first machine learning model by a first device as a result of processing audio data, by the first device, using the first machine learning model;
receiving, from a second device, first audio data corresponding to operation of the first machine learning model by the second device;
processing the first audio data using a second machine learning model to determine the first audio data corresponds to incorrect processing by the second device using the first machine learning model; and
processing the first audio data and the first parameter data to determine an updated first machine learning model.
6. The computer-implemented method of claim 5 , further comprising:
processing the first audio data using a third machine learning model to determine label data corresponding to the first audio data,
wherein determination of the updated first machine learning model further comprises processing the label data.
7. The computer-implemented method of claim 6 , further comprising:
performing automatic speech recognition (ASR) processing using the first audio data to determine ASR data; and
performing natural language understanding (NLU) processing using the ASR data to determine NLU data,
wherein determination of the label data further comprises processing the ASR data and the NLU data.
8. The computer-implemented method of claim 5 , further comprising:
receiving first metadata corresponding to the first parameter data; and
receiving second metadata corresponding to the first audio data,
wherein determination of the updated first machine learning model further comprises processing the first metadata and the second metadata.
9. The computer-implemented method of claim 8 , wherein the second metadata comprises an indicator of a wakeword detected by the second device.
10. The computer-implemented method of claim 5 , further comprising:
receiving second parameter data corresponding to adjustment of at least one parameter of the first machine learning model by a third device as a result of processing audio data, by the third device, using the first machine learning model;
storing the first parameter data and the second parameter data; and
determining stored parameter data satisfies a condition corresponding to an amount of the stored parameter data,
wherein determination of the updated first machine learning model is performed in response to the stored parameter data satisfying the condition, and
wherein determination of the updated first machine learning model further comprises processing the second parameter data.
11. The computer-implemented method of claim 5 , further comprising:
receiving, from a third device, second parameter data;
determining the first parameter data is associated with a first characteristic;
determining the first audio data is associated with the first characteristic; and
determining the second parameter data is associated with a second characteristic different from the first characteristic,
wherein determination of the updated first machine learning model is performed without involving the second parameter data in response to the second parameter data being associated with the second characteristic.
12. The computer-implemented method of claim 5 , further comprising:
determining the updated first machine learning model satisfies a difference condition with respect to the first machine learning model; and
in response to satisfaction of the difference condition, sending, to the first device and the second device, first data corresponding to the updated first machine learning model.
13. A system comprising:
at least one processor; and
at least one memory comprising instructions that, when executed by the at least one processor, cause the system to:
receive first parameter data corresponding to adjustment of at least one parameter of a first machine learning model by a first device as a result of processing audio data, by the first device, using the first machine learning model;
receive, from a second device, first audio data corresponding to operation of the first machine learning model by the second device;
process the first audio data using a second machine learning model to determine the first audio data corresponds to incorrect processing by the second device using the first machine learning model; and
process the first audio data and the first parameter data to determine an updated first machine learning model.
14. The system of claim 13 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
process the first audio data using a third machine learning model to determine label data corresponding to the first audio data,
wherein determination of the updated first machine learning model further comprises processing the label data.
15. The system of claim 14 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
perform automatic speech recognition (ASR) processing using the first audio data to determine ASR data; and
perform natural language understanding (NLU) processing using the ASR data to determine NLU data,
wherein determination of the label data further comprises processing the ASR data and the NLU data.
16. The system of claim 13 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
receive first metadata corresponding to the first parameter data; and
receive second metadata corresponding to the first audio data,
wherein determination of the updated first machine learning model further comprises processing the first metadata and the second metadata.
17. The system of claim 16 , wherein the second metadata comprises an indicator of a wakeword detected by the second device.
18. The system of claim 13 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
receive second parameter data corresponding to adjustment of at least one parameter of the first machine learning model by a third device as a result of processing audio data, by the third device, using the first machine learning model;
store the first parameter data and the second parameter data; and
determine stored parameter data satisfies a condition corresponding to an amount of the stored parameter data,
wherein determination of the updated first machine learning model is performed in response to the stored parameter data satisfying the condition, and
wherein determination of the updated first machine learning model further comprises processing the second parameter data.
19. The system of claim 13 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
receive, from a third device, second parameter data;
determine the first parameter data is associated with a first characteristic;
determine the first audio data is associated with the first characteristic; and
determine the second parameter data is associated with a second characteristic different from the first characteristic,
wherein determination of the updated first machine learning model is performed without involving the second parameter data in response to the second parameter data being associated with the second characteristic.
20. The system of claim 13 , wherein the at least one memory further comprises instructions that, when executed by the at least one processor, further cause the system to:
determine the updated first machine learning model satisfies a difference condition with respect to the first machine learning model; and
in response to satisfaction of the difference condition, send, to the first device and the second device, first data corresponding to the updated first machine learning model.