Removing bias from automatic speech recognition models using internal language model estimates
Bias may be removed from automatic speech recognition model predictions using internal language model estimates. Audio data may be received for speech recognition. The audio data may be processed both through an automatic speech recognition model to produce original word token predictions and masked in different portions of the audio data to produce other word token predictions for the masked audio. A comparison of the original word token predictions and the other word token predictions may provide an estimate of an internal language model for the automatic speech recognition model. This estimate can be used to modify the original word token predictions to remove the lexical bias and produce a speech prediction.
1. A system, comprising:
at least one processor; and
a memory, storing program instructions that when executed by the at least one processor, cause the at least one processor to implement a speech recognition system, configured to:
receive audio data to recognize speech in the audio data;
cause the audio data to be processed through a machine learning model trained for speech recognition that outputs original token predictions corresponding to respective words recognized in the audio data by the machine learning model;
generate a plurality of masked versions of the audio data, wherein individual ones of plurality of masked version of the audio data comprise different respective masked portions of the audio data;
cause the plurality of masked versions of the audio data to be processed through the machine learning model to output internal language model estimation token predictions corresponding to respective words recognized in the plurality of masked versions of the audio data by the machine learning model;
compare the internal language model estimation token predictions with the original token predictions to determine modifications to one or more of the original token predictions according to differences between the internal language model estimation token predictions with the original token predictions that are above difference threshold;
apply the modifications to the one or more original token predictions; and
generate a speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions.
2. The system of claim 1 , wherein to generate the speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions, the speech recognition system is configured to apply a beam search technique using the original token predictions fused with an external machine learning language model.
3. The system of claim 1 , wherein the different respective masked portions correspond to different ones of equal partitions of the audio data.
4. The system of claim 1 , wherein the speech recognition system is implemented as part of a natural language processing service offered by a provider network and wherein the speech prediction is provided to one or more other resources of the natural language processing service to perform a natural language processing task on the speech prediction.
5. A method, comprising:
receiving, at a speech recognition system, audio data to recognize speech in the audio data;
processing, by the speech recognition system, the audio data through a machine learning model trained for speech recognition that outputs original token predictions corresponding to respective words recognized in the audio data by the machine learning model;
generating, by the speech recognition system, a plurality of masked versions of the audio data, wherein individual ones of plurality of masked version of the audio data comprise different respective masked portions of the audio data;
processing, by the speech recognition system, the plurality of masked versions of the audio data through the machine learning model to output internal language model estimation token predictions corresponding to respective words recognized in the plurality of masked versions of the audio data by the machine learning model;
comparing, by the speech recognition system, the internal language model estimation token predictions with the original token predictions to apply modifications to one or more of the original token predictions according to differences between the internal language model estimation token predictions with the original token predictions that are above difference threshold; and
generating, by the speech recognition system, a speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions.
6. The method of claim 5 , wherein the machine learning model is a non-autoregressive model.
7. The method of claim 5 , wherein generating the speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions comprises applying a beam search technique using the original token predictions fused with an external machine learning language model.
8. The method of claim 5 , wherein generating the plurality of masked versions of the audio data is performed according to a masking scheme to select the different respective masked portions, and wherein the masking scheme is specified according to a request received via an interface of the speech recognition system.
9. The method of claim 5 , wherein the different respective masked portions correspond to different ones of equal partitions of the audio data.
10. The method of claim 5 , wherein the machine learning model is a deep neural network trained according to Connectionist Temporal Classification (CTC) technique, and wherein a blank token of the original token predictions output by the machine learning model is not modified according to the comparing.
11. The method of claim 5 , wherein the difference threshold is specified according to a request received via an interface of the speech recognition system.
12. The method of claim 5 , wherein the speech prediction is one of a plurality of possible speech predictions generated using the original token predictions including the modified one or more original token predictions.
13. The method of claim 5 , wherein the speech recognition system is implemented as part of a natural language processing service offered by a provider network and wherein the speech prediction is provided to one or more other resources of the natural language processing service to perform a natural language processing task on the speech prediction.
14. One or more non-transitory, computer-readable storage media, storing program instructions that when executed on or across one or more computing devices cause the one or more computing devices to implement:
receiving audio data to recognize speech in the audio data;
causing the audio data to be processed through a machine learning model trained for speech recognition that outputs original token predictions corresponding to respective words recognized in the audio data by the machine learning model;
generating a plurality of masked versions of the audio data, wherein individual ones of plurality of masked version of the audio data comprise different respective masked portions of the audio data;
causing the plurality of masked versions of the audio data to be processed through the machine learning model to output internal language model estimation token predictions corresponding to respective words recognized in the plurality of masked versions of the audio data by the machine learning model;
comparing the internal language model estimation token predictions with the original token predictions to apply modifications to one or more of the original token predictions according to differences between the internal language model estimation token predictions with the original token predictions that are above difference threshold; and
generating a speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions.
15. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the machine learning model is an autoregressive model.
16. The one or more non-transitory, computer-readable storage media of claim 14 , wherein, in generating the speech prediction in the audio data according to the original token predictions including the modified one or more original token predictions, the program instructions cause the one or more computing devices to implement applying a beam search technique using the original token predictions.
17. The one or more non-transitory, computer-readable storage media of claim 14 , wherein generating the plurality of masked versions of the audio data is performed according to a masking scheme to select the different respective masked portions, and wherein the masking scheme is specified according to a request received via an interface of a speech recognition system.
18. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the different respective masked portions correspond to different words in the audio data.
19. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the difference threshold is specified according to a request received via an interface of the speech recognition system.
20. The one or more non-transitory, computer-readable storage media of claim 14 , wherein the one or more computing devices are implemented as part of a natural language processing service offered by a provider network and wherein the speech prediction is provided to one or more other resources of the natural language processing service to perform a natural language processing task on the speech prediction.