Characterizing accuracy of ensemble models for automatic speech recognition by determining a predetermined number of multiple ASR engines based on their historical performance

1. A method for analyzing a transcription of a recording, comprising:

generating features representing transcriptions produced by multiple automatic speech recognition (ASR) engines from voice activity in a recording;

generating a best transcription of the recording produced by an ensemble model from the transcriptions;

applying a machine learning model to the features to produce a score representing an accuracy of the best transcription, wherein applying the machine learning model to the features to produce the score representing the accuracy of the best transcription comprises:

determining a predetermined number of the multiple ASR engines having historical performance in generating transcripts that are better than the historical performance of a remaining number of the multiple ASR engines under predetermined conditions as contributor ASR engines, the pre-determined conditions comprising a selected metadata associated with the recording,

determining a remaining number of the multiple ASR engines as selector ASR engines, and

applying the machine learning model to produce the score representing error rates between the best transcription and the transcriptions from the selector ASR engines; and

storing the score in association with the best transcription.

2. The method of claim 1 , wherein generating the features representing the transcriptions produced by the multiple ASR engines from voice activity in the recording and the best transcription of the recording produced by the ensemble model from the transcriptions comprises:

generating a first set of features from the transcriptions;

generating a second set of features from pairwise comparisons of the transcriptions;

generating a third set of features from the best transcription; and

generating a fourth set of features from the recording.

3. The method of claim 2 , wherein the fourth set of features comprises a duration of the voice activity, a position of the voice activity in the recording, and an audio feature.

4. The method of claim 3 , wherein the audio feature comprises at least one of a mel-frequency cepstral coefficient (MFCC), a perceptual linear prediction (PLP), a root mean square (RMS), a zero crossing rate, a spectral flux, a spectral energy, a chroma vector, and a chroma deviation.

5. The method of claim 2 , wherein the first set of features comprises a length of a transcription, a confidence in the transcription, and a letters per second associated with the transcription.

6. The method of claim 2 , wherein the second set of features comprises a word error rate between two transcriptions, an average word error rate across all pairs of transcriptions, a difference in length between the two transcriptions, and an average difference in length across all pairs of transcriptions.

7. The method of claim 2 , wherein the third set of features comprises a first feature representing a pairwise comparison of the best transcription and each of the transcriptions and a second feature representing an attribute of the best transcription.

8. The method of claim 2 , wherein the second set of features comprises a fixed-size encoding of per-character differences between two transcriptions.

9. The method of claim 1 , further comprising:

applying one or more thresholds to the score to characterize the accuracy of the best transcription; and

determining a candidacy of the recording for human transcription based on the characterized accuracy of the best transcription.

10. The method of claim 9 , further comprising:

generating training data for the ensemble model from the best transcription and the human transcription; and

updating parameters of the ensemble model based on the training data.

11. The method of claim 9 , wherein determining the candidacy of the recording for the human transcription based on the characterized accuracy of the best transcription comprises:

identifying the recording as a candidate for the human transcription when the score falls between a first threshold for a high error rate and a second threshold for a low error rate.

12. The method of claim 1 , wherein the machine learning model comprises an artificial neural network.

13. A non-transitory computer readable medium storing instructions that, when executed by a processor, cause the processor to perform the steps of:

generating features representing transcriptions produced by multiple automatic speech recognition (ASR) engines from voice activity in the recording and a best transcription of the recording produced by an ensemble model from the transcriptions;

applying a machine learning model to the features to produce a score representing an accuracy of the best transcription, wherein applying the machine learning model to the features to produce the score representing the accuracy of the best transcription comprises:

determining a predetermined number of the multiple ASR engines having historical performance in generating transcripts that are better than the historical performance of a remaining number of the multiple ASR engines under predetermined conditions as contributor ASR engines, the pre-determined conditions comprising a selected metadata associated with the recording,

determining a remaining number of the multiple ASR engines as selector ASR engines, and

applying the machine learning model to produce the score representing error rates between the best transcription and the transcriptions from the selector ASR engines; and

storing the score in association with the best transcription.

14. The non-transitory computer readable medium of claim 13 , wherein generating the features representing the transcriptions produced by the multiple ASR engines from voice activity in the recording and the best transcription of the recording produced by the ensemble model from the transcriptions comprises:

generating a first set of features from the transcriptions;

generating a second set of features from pairwise comparisons of the transcriptions;

generating a third set of features from the best transcription; and

generating a fourth set of features from the recording.

15. The non-transitory computer readable medium of claim 14 , wherein the first set of features and third set of features comprise a length of a transcription, a confidence in the transcription, and a letters per second associated with the transcription.

16. The non-transitory computer readable medium of claim 14 , wherein the fourth set of features comprises a duration of the voice activity, a position of the voice activity in the recording, and an audio feature.

17. The non-transitory computer readable medium of claim 13 , wherein the steps further comprise:

applying one or more thresholds to the score to characterize the accuracy of the best transcription; and

determining a candidacy of the recording for human transcription based on the characterized accuracy of the best transcription.

18. The non-transitory computer readable medium of claim 17 , wherein the steps further comprise:

generating training data for the ensemble model from the best transcription and the human transcription; and

updating parameters of the ensemble model based on the training data.

19. The non-transitory computer readable medium of claim 14 , wherein the second set of features and third set of features comprise a word error rate between two transcriptions, an average word error rate across all pairs of transcriptions, a difference in length between the two transcriptions, and an average difference in length across all pairs of transcriptions.

20. A system, comprising:

a memory that stores instructions; and

a processor that is coupled to the memory and, when executing the instructions, is configured to:

generate features representing transcriptions produced by multiple automatic speech recognition (ASR) engines from voice activity in the recording and a best transcription of the recording produced by an ensemble model from the transcriptions;

apply a machine learning model to the features to produce a score representing an accuracy of the best transcription, wherein applying the machine learning model to the features to produce the score representing the accuracy of the best transcription comprises:

determining a predetermined number of the multiple ASR engines having historical performance in generating transcripts that are better than the historical performance of a remaining number of the multiple ASR engines under predetermined conditions as contributor ASR engines, the pre-determined conditions comprising a selected metadata associated with the recording,

determining a remaining number of the multiple ASR engines as selector ASR engines, and

applying the machine learning model to produce the score representing error rates between the best transcription and the transcriptions from the selector ASR engines; and

store the score in association with the best transcription.