System and method for performing automatic speech recognition system parameter adjustment via machine learning

1. A method comprising:

receiving first speech at an automatic speech recognition system;

estimating, from the first speech via a model trained on metadata to output parameters associated with a chosen acoustic environment, a first set of parameters associated with the chosen acoustic environment;

receiving second speech at the automatic speech recognition system;

extracting features from the second speech;

refining the estimated first set of parameters based on the extracted features from the second speech;

applying, by the automatic speech recognition system, the refined estimated first set of parameters to recognize the first speech and the second speech to yield text; and

outputting the text from the automatic speech recognition system.

2. The method of claim 1 , wherein the chosen acoustic environment is selected from a plurality of different acoustic environments.

3. The method of claim 1 , wherein the first set of parameters comprises one or more of a word insertion penalty, a silence prior, a word penalty, a beam pruning width, a language model scale, an acoustic model scale, a duration model scale, other search pruning control parameters, and a language model interpolation vector.

4. The method of claim 1 , further comprising, prior to estimating the first set of parameters, applying initial parameters for recognizing initial speech received at the automatic speech recognition system, wherein the first set of parameters replace the initial parameters.

5. The method of claim 1 , wherein the model utilizes one or more of a signal-to-noise ratio estimate, reverberation time estimate, a short-term window frequency analysis, a mel-scale frequency cepstral analysis, time-domain signal audio signal directly and the metadata to estimate the first set of parameters.

6. The method of claim 1 , wherein the metadata includes non-audio data.

7. The method of claim 1 , wherein the model is separate from the automatic speech recognition system and comprises one of a feedforward neural network, unidirectional or bidirectional recurrent neural network, a convolutional neural network or a support vector machine model.

8. The method of claim 1 , wherein the metadata is incorporated into the model via one-hot encoding or embedding.

9. The method of claim 1 , wherein estimating the first set of parameters and applying the refined estimated first set of parameters to recognize the first speech and the second speech to yield the text is performed in a batch mode.

10. The method of claim 9 , wherein the applying of the refined estimated first set of parameters to recognize the first speech and the second speech to yield the text is performed in either a delayed decoding pass by the automatic speech recognition system, or in a rescoring that rescores result options from a first speech recognition pass given estimated hyper parameters.

11. The method of claim 10 , wherein the estimated hyper parameters estimated on one utterance are only applied in decoding of a respectively next utterance.

12. The method of claim 1 , wherein estimating the first set of parameters useful for performing automatic speech recognition yields (1) the first set of parameters directly as target layer outputs or (2) the first set of parameters as a predefined parameter configuration chosen from a group of predefined parameter configurations.

13. An automatic speech recognition system comprising:

a processor; and

a non-transitory computer-readable storage medium having instructions stored which, when executed by the processor, cause the processor to perform operations comprising:

estimating, via a model, a first set of parameters associated with a chosen acoustic environment, the first set of parameters being useful for performing automatic speech recognition, wherein the estimating is performed during speech recognition of first speech received by the automatic speech recognition system;

generating a list of recognition results using the first set of parameters;

receiving second speech at the automatic speech recognition system;

estimating, via the model, a second set of parameters based on processing the second speech; and

applying, by the automatic speech recognition system, the second set of parameters to rescore the list of recognition results.

14. The automatic speech recognition system of claim 13 , wherein the chosen acoustic environment is selected from a plurality of different acoustic environments.

15. The automatic speech recognition system of claim 13 , wherein the first set of parameters comprises one or more of a word insertion penalty, a silence prior, a word penalty, a beam pruning width, a language model scale, an acoustic model scale, a duration model scale, other search pruning control parameters, or a language model interpolation vector.

16. The automatic speech recognition system of claim 13 , wherein the non-transitory computer-readable storage medium stores further instructions which, when executed by the processor, cause the processor to perform further operations comprising:

prior to estimating the first set of parameters, applying initial parameters for recognizing initial speech received at the automatic speech recognition system, wherein the first set of parameters replace the initial parameters.

17. The automatic speech recognition system of claim 13 , wherein the model utilizes one or more of a signal-to-noise ratio estimate, reverberation time estimate, a short-term window frequency analysis, a mel-scale frequency cepstral analysis, time-domain signal audio signal directly and the metadata to estimate the first set of parameters.

18. The automatic speech recognition system of claim 13 , wherein the model is trained on metadata to output the first set of parameters, wherein the metadata comprises non-audio data.

19. The automatic speech recognition system of claim 13 , wherein estimating the first set of parameters useful for performing automatic speech recognition yields (1) the first set of parameters directly as target layer outputs or (2) the first set of parameters as a predefined parameter configuration chosen from a group of predefined parameter configurations.

20. A method comprising:

receiving speech at an automatic speech recognition system;

extracting first features from a first portion of the speech;

inputting the extracted first features into a machine learning model that estimates a first set of parameters useful for performing automatic speech recognition by the automatic speech recognition system;

applying the first set of parameters to the automatic speech recognition system to yield a tuned automatic speech recognition system;

extracting second features from a second portion of the speech;

inputting the extracted second features into the machine learning model that refines the estimated first set of parameters;

recognizing, by applying the refined estimated first set of parameters, the first portion of speech and the second portion of speech by the tuned automatic speech recognition system to yield text; and

outputting the text from the tuned automatic speech recognition system.