Method and apparatus for predicting word accuracy in automatic speech recognition systems

1. A method for predicting a word accuracy, comprising:

obtaining, by a processor, an utterance in speech data, wherein the utterance comprises an actual word string;

processing, by the processor, the utterance for generating an interpretation of the actual word string;

processing, by the processor, the utterance to identify an utterance frame; and

calculating, by the processor, a prediction of a word accuracy associated with the interpretation based on a stationary signal-to-noise ratio and a non-stationary signal-to-noise ratio, wherein the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio are determined according to a frame energy associated with the utterance frame, and wherein the calculating comprises:

computing the stationary signal-to-noise ratio for the utterance;

computing the non-stationary signal-to-noise ratio for the utterance; and

computing the prediction of the word accuracy associated with the interpretation using the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio.

2. The method of claim 1 , wherein the obtaining the utterance in speech data, the processing the utterance to generate the interpretation of the actual word string, the processing the utterance to identify the utterance frame, and the calculating the prediction of the word accuracy are repeated for an additional utterance.

3. The method of claim 1 , wherein the processing the utterance to generate the interpretation of the actual word string, comprises:

obtaining an acoustic model;

obtaining a language model; and

applying the acoustic model and the language model to the utterance for generating the interpretation.

4. The method of claim 1 , wherein the processing the utterance to identify the utterance frame comprises:

aligning the utterance to the interpretation;

partitioning the utterance into the utterance frame; and

classifying the utterance frame as one of: a speech frame or a silence frame.

5. The method of claim 4 , wherein the aligning is performed according to a forced alignment algorithm.

6. The method of claim 4 , wherein the classifying is performed according to a speech-silence decoding algorithm.

7. The method of claim 1 , wherein the computing the stationary signal-to-noise ratio comprises:

computing an average signal power for the utterance;

computing an average noise power for the utterance; and

calculating the stationary signal-to-noise ratio using the average signal power and the average noise power.

8. The method of claim 7 , wherein the average signal power and the average noise power are computed using the frame energy associated with the utterance frame.

9. The method of claim 1 , wherein the computing the non-stationary signal-to-noise ratio comprises:

computing an average signal power for the utterance;

determining a noise power for each of a plurality of utterance frames of the utterance classified as silence frames;

computing a noise signal-to-noise ratio for each of the plurality of utterance frames, wherein the noise signal-to-noise ratio for one of the plurality of utterance frames is computed using the average signal power and the noise power of the one of the plurality of utterance frames; and

computing a standard deviation of the noise signal-to-noise ratios for the plurality of utterance frames.

10. The method of claim 9 , wherein the average signal power and the noise power are computed using the frame energy associated with each of the plurality of utterance frames.

11. The method of claim 1 , wherein the computing the prediction of the word accuracy associated with the interpretation is performed according to a linear least square algorithm.

12. The method of claim 1 , wherein the computing the prediction of the word accuracy associated with the interpretation is performed using a confidence score.

13. The method of claim 1 , wherein computing the frame energy comprises:

sampling the utterance frame to produce an utterance frame sample;

generating a squared utterance frame sample by squaring the utterance frame sample;

computing a summation of the squared utterance frame sample; and

computing a logarithm of the summation.

14. A non-transitory computer readable medium storing a software program, that when executed by a computer, causes the computer to perform operations, the operations comprising:

obtaining an utterance in speech data, wherein the utterance comprises an actual word string;

processing the utterance for generating an interpretation of the actual word string;

processing the utterance to identify an utterance frame; and

calculating a prediction of a word accuracy associated with the interpretation based on a stationary signal-to-noise ratio and a non-stationary signal-to-noise ratio, wherein the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio are determined according to a frame energy associated with the utterance frame, and wherein the calculating comprises:

computing the stationary signal-to-noise ratio for the utterance;

computing the non-stationary signal-to-noise ratio for the utterance; and

computing the prediction of the word accuracy associated with the interpretation using the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio.

15. The non-transitory computer readable medium of claim 14 , wherein the processing the utterance to generate the interpretation of the actual word string, comprises:

obtaining an acoustic model;

obtaining a language model; and

applying the acoustic model and the language model to the utterance for generating a predicted word string.

16. The non-transitory computer readable medium of claim 14 , wherein the processing the utterance to identify the utterance frame, comprises:

aligning the utterance to the interpretation;

partitioning the utterance into the utterance frame; and

classifying the utterance frame as one of: a speech frame or a silence frame.

17. The non-transitory computer readable medium of claim 14 , wherein the computing the non-stationary signal-to-noise ratio, comprises:

computing an average signal power for the utterance;

determining a noise power for each of a plurality of utterance frames of the utterance classified as silence frames;

computing a noise signal-to-noise ratio for each of the plurality of utterance frames, wherein the noise signal-to-noise ratio for one of the plurality of utterance frames is computed using the average signal power and the noise power of the one of the plurality of utterance frames; and

computing a standard deviation of the noise signal-to-noise ratios for the plurality of utterance frames.

18. An apparatus for predicting a word accuracy, comprising:

a processor; and

a computer-readable medium in communication with the processor, storing a plurality of instructions which, when executed by the processor, cause the processor to perform operations, the operations comprising:

obtaining an utterance in speech data, wherein the utterance comprises an actual word string;

processing the utterance for generating an interpretation of the actual word string;

processing the utterance to identify an utterance frame; and

calculating a prediction of a word accuracy associated with the interpretation based on a stationary signal-to-noise ratio and a non-stationary signal-to-noise ratio, wherein the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio are determined according to a frame energy associated with the utterance frame, and wherein the calculating the prediction of the word accuracy associated with the interpretation comprises:

computing the stationary signal-to-noise ratio for the utterance;

computing the non-stationary signal-to-noise ratio for the utterance; and

computing the prediction of the word accuracy associated with the interpretation using the stationary signal-to-noise ratio and the non-stationary signal-to-noise ratio.

19. The apparatus of claim 18 , wherein the obtaining the utterance in speech data, the processing the utterance to generate the interpretation of the actual word string, the processing the utterance to identify the utterance frame, and the calculating the prediction of the word accuracy are repeated for an additional utterance.

20. The apparatus of claim 18 , wherein the processing the utterance to generate the interpretation of the actual word string, comprises:

obtaining an acoustic model;

obtaining a language model; and

applying the acoustic model and the language model to the utterance for generating the interpretation.