Acoustically informed pruning for language modeling

1. A computer-implemented method comprising:

selecting a candidate portion of a language model to evaluate for pruning;

obtaining an entropy score representing information loss that would result from pruning the candidate portion of the language model;

obtaining an acoustic score representing acoustic confusability of one or more words modeled by the candidate portion of the language model;

evaluating whether to prune the candidate portion of the language model using the entropy score and the acoustic score;

generating a pruned language model based on pruning the candidate portion from the language model; and

generating, by a speech recognizer using the pruned language model, a transcription of an utterance.

2. The method of claim 1 , wherein the language model includes sequences of contiguous items and the candidate portion includes a sequence of contiguous items.

3. The method of claim 2 , wherein the candidate portion includes a homophone and the language model includes second contiguous sequence of items that includes a corresponding homophone.

4. The method of claim 3 , wherein obtaining the acoustic score includes determining whether the candidate portion includes a homophone.

5. The method of claim 1 , wherein evaluating includes:

obtaining a pruning threshold;

adjusting the pruning threshold based on the acoustic score;

comparing the entropy score to the adjusted pruning threshold; and

determining whether to prune the candidate portion based on the comparison.

6. The method of claim 1 , wherein evaluating includes:

obtaining a pruning threshold;

adjusting the entropy score based on the acoustic score;

comparing the adjusted entropy score to the pruning threshold; and

determining whether to prune the candidate portion based on the comparison.

7. A system comprising:

one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:

selecting a candidate portion of a language model to evaluate for pruning;

obtaining an entropy score representing information loss that would result from pruning the candidate portion of the language model;

obtaining an acoustic score representing acoustic confusability of one or more words modeled by the candidate portion of the language model;

evaluating whether to prune the candidate portion of the language model using the entropy score and the acoustic score; and

generating a pruned language model based on pruning the candidate portion from the language model; and

generating, by a speech recognizer using the pruned language model, a transcription of an utterance.

8. The system of claim 7 , wherein the language model includes sequences of contiguous items and the candidate portion includes a sequence of contiguous items.

9. The system of claim 8 , wherein the candidate portion includes a homophone and the language model includes second contiguous sequence of items that includes a corresponding homophone.

10. The system of claim 9 , wherein obtaining the acoustic score includes determining whether the candidate portion includes a homophone.

11. The system of claim 7 , wherein evaluating includes:

obtaining a pruning threshold;

adjusting the pruning threshold based on the acoustic score;

comparing the entropy score to the adjusted pruning threshold; and

determining whether to prune the candidate portion based on the comparison.

12. The system of claim 7 , wherein evaluating includes:

obtaining a pruning threshold;

adjusting the entropy score based on the acoustic score;

comparing the adjusted entropy score to the pruning threshold; and

determining whether to prune the candidate portion based on the comparison.

13. A non-transitory computer-readable medium storing software comprising instructions executable by one or more computers which, upon such execution, cause the one or more computers to perform operations comprising:

selecting a candidate portion of a language model to evaluate for pruning;

obtaining an entropy score representing information loss that would result from pruning the candidate portion of the language model;

obtaining an acoustic score representing acoustic confusability of one or more words modeled by the candidate portion of the language model;

evaluating whether to prune the candidate portion of the language model using the entropy score and the acoustic score;

generating a pruned language model based on pruning the candidate portion from the language model; and

generating, by a speech recognizer using the pruned language model, a transcription of an utterance.

14. The medium of claim 13 , wherein the language model includes sequences of contiguous items and the candidate portion includes a sequence of contiguous items.

15. The medium of claim 14 , wherein the candidate portion includes a homophone and the language model includes second contiguous sequence of items that includes a corresponding homophone.

16. The medium of claim 15 , wherein obtaining the acoustic score includes determining whether the candidate portion includes a homophone.

17. The medium of claim 13 , wherein evaluating includes:

obtaining a pruning threshold;

adjusting the pruning threshold based on the acoustic score;

comparing the entropy score to the adjusted pruning threshold; and

determining whether to prune the candidate portion based on the comparison.