Using context information with end-to-end models for speech recognition

1. A method comprising:

receiving, at data processing hardware, audio data encoding an utterance;

processing, by the data processing hardware, using a speech recognition model, the audio data to generate speech recognition scores for speech elements;

determining, by the data processing hardware, that a preliminary transcription for the utterance comprises a word that represents a prefix element, the prefix element indicating that a next element corresponds to a context for the utterance;

selecting, by the data processing hardware, a first contextual finite-state transducer (FST) from a plurality of contextual FSTs based on a particular context corresponding to the first contextual FST matching the context for the utterance indicated by the word of the preliminary transcription determined to represent the prefix element of the utterance, wherein each contextual FST in the plurality of contextual FSTs corresponds to a respective different particular context for a same user that spoke the utterance;

determining, by the data processing hardware, using the first contextual FST, context scores for the speech elements;

executing, by the data processing hardware, using the speech recognition scores and the context scores, a beam search decoding process to determine one or more candidate transcriptions for the utterance; and

selecting, by the data processing hardware, a transcription for the utterance from the one or more candidate transcriptions.

2. The method of claim 1 , wherein, during execution of the beam search decoding process, the context scores are configured to adjust a likelihood of the one or more candidate transcriptions before pruning any of the one or more candidate transcriptions from evaluation.

3. The method of claim 1 , wherein executing the beam search decoding process comprises using the context scores to prune paths through a speech recognition lattice to determine the one or more candidate transcriptions for the utterance.

4. The method of claim 1 , further comprising, prior to receiving the audio data encoding the utterance:

generating, by the data processing hardware, the plurality of contextual FSTs to each represent a different set of words or phrases in a personalized data collection of the same user that spoke the utterance; and

storing, by the data processing hardware, the plurality of contextual FSTs in memory hardware in communication with the data processing hardware.

5. The method of claim 4 , wherein the personalized data collection comprises a contacts list for the same user.

6. The method of claim 4 , wherein the personalized data collection comprises a media library for the same user.

7. The method of claim 4 , wherein the personalized data collection comprises a list of applications installed on a user device associated with the same user.

8. The method of claim 4 , further comprising, for each of at least one contextual FST in the plurality of contextual FSTs:

generating, by the data processing hardware, a corresponding prefix FST comprising a set of one or more prefixes each corresponding to the respective different particular context of the corresponding contextual FST; and

storing, by the data processing hardware, the corresponding prefix FST generated for the at least one contextual FST in the plurality of contextual FSTs.

9. The method of claim 1 , wherein the data processing hardware:

resides on a user device associated with the same user that spoke the utterance; and

executes the speech recognition model.

10. The method of claim 1 , wherein the speech recognition model comprises an end-to-end speech recognition model.

11. The method of claim 10 , wherein the end-to-end speech recognition model comprises a recurrent neural network-transducer (RNN-T).

12. The method of claim 1 , wherein the plurality of contextual FSTs represent contextual terms using elements representing subword units.

13. The method of claim 1 , wherein the plurality of contextual FSTs comprise:

transition weights configured to bias transitions between subword units of a contextual term; and

backoff arcs having offsetting weights configured to undo the biasing effect of the transition weight.

14. The method of claim 1 , wherein the speech elements comprise wordpieces or graphemes.

15. A system comprising:

data processing hardware; and

memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:

receiving audio data encoding an utterance;

processing, using a speech recognition model, the audio data to generate speech recognition scores for speech elements;

determining that a preliminary transcription for the utterance comprises a word that represents a prefix element, the prefix element indicating that a next element corresponds to a context for the utterance;

selecting a first contextual finite-state transducer (FST) from a plurality of contextual FSTs based on a particular context corresponding to the first contextual FST matching the context for the utterance indicated by the word of the preliminary transcription determined to represent the prefix element of the utterance, wherein each contextual FST in the plurality of contextual FSTs corresponds to a respective different particular context for a same user that spoke the utterance;

determining, using the first FST, context scores for the speech elements;

executing, using the speech recognition scores and the context scores, a beam search decoding process to determine one or more candidate transcriptions for the utterance; and

selecting a transcription for the utterance from the one or more candidate transcriptions.

16. The system of claim 15 , wherein, during execution of the beam search decoding process, the context scores are configured to adjust a likelihood of the one or more candidate transcriptions before pruning any of the one or more candidate transcriptions from evaluation.

17. The system of claim 15 , wherein executing the beam search decoding process comprises using the context scores to prune paths through a speech recognition lattice to determine the one or more candidate transcriptions for the utterance.

18. The system of claim 15 , wherein the operations further comprise, prior to receiving the audio data encoding the utterance:

generating the plurality of contextual FSTs to each represent a different set of words or phrases in a personalized data collection of the same user that spoke the utterance; and

storing the plurality of contextual FSTs in the memory hardware in communication with the data processing hardware.

19. The system of claim 18 , wherein the personalized data collection comprises a contacts list for the same user.

20. The system of claim 18 , wherein the personalized data collection comprises a media library for the same user.

21. The system of claim 18 , wherein the personalized data collection comprises a list of applications installed on a user device associated with the same user.

22. The system of claim 18 , wherein the operations further comprise, for each of at least one contextual FST in the plurality of contextual FSTs:

generating a corresponding prefix FST comprising a set of one or more prefixes each corresponding to the respective different particular context of the corresponding contextual FST; and

storing the corresponding prefix FST generated for the at least one contextual FST in the plurality of contextual FSTs.

23. The system of claim 15 , wherein the data processing hardware:

resides on a user device associated with the same user that spoke the utterance; and

executes the speech recognition model.

24. The system of claim 15 , wherein the speech recognition model comprises an end-to-end speech recognition model.

25. The system of claim 24 , wherein the end-to-end speech recognition model comprises a recurrent neural network-transducer (RNN-T).

26. The system of claim 15 , wherein the plurality of contextual FSTs represent contextual terms using elements representing subword units.

27. The system of claim 26 , wherein the plurality of contextual FSTs comprise:

transition weights configured to bias transitions between subword units of a contextual term; and

backoff arcs having offsetting weights configured to undo the biasing effect of the transition weight.

28. The system of claim 15 , wherein the speech elements comprise wordpieces or graphemes.