Lookup-table recurrent language model
A computer-implemented method includes receiving audio data that corresponds to an utterance spoken by a user and captured by a user device. The method also includes processing the audio data to determine a candidate transcription that includes a sequence of tokens for the spoken utterance. Tor each token in the sequence of tokens, the method includes determining a token embedding for corresponding token, determining a n-gram token embedding for a previous sequence of n-gram tokens, and concatenating the token embedding and the n-gram token embedding to generate a concatenated output for the corresponding token. The method also includes rescoring the candidate transcription for the spoken utterance by processing the concatenated output generated for each corresponding token in the sequence of tokens.
1. A computer-implemented method when executed on data processing hardware causes the data processing hardware to perform operations, the operations comprising:
receiving audio data corresponding to an utterance spoken by a user and captured by a user device;
processing, using a speech recognizer, the audio data to determine a candidate transcription for the spoken utterance, the candidate transcription comprises a sequence of tokens;
for each corresponding token in the sequence of tokens subsequent to an initial token in the sequence of tokens:
determining, using a first embedding table, a token embedding for the corresponding token, the token embedding determined independent from each token in the sequence of tokens that precedes the corresponding token in the sequence of tokens;
determining, using a second embedding table, a n-gram token embedding for a previous sequence of n-gram tokens, based on each token in the sequence of tokens that precedes the corresponding token in the sequence of tokens; and
concatenating the token embedding and the n-gram token embedding to generate a concatenated output for the corresponding token; and
rescoring, using an external language model, the candidate transcription for the spoken utterance by processing the concatenated output generated for each corresponding token in the sequence of tokens.
2. The method of claim 1 , wherein the external language model comprises a recurrent neural network language model.
3. The method of claim 1 , wherein the external language model integrates with the speech recognizer by Hybrid Autoregressive Transducer (HAT) factorizing the speech recognizer.
4. The method of claim 1 , wherein the speech recognizer comprises a conformer audio encoder and a recurrent neural network-transducer decoder.
5. The method of claim 1 , wherein the speech recognizer comprises a transformer audio encoder and a recurrent neural network-transducer decoder.
6. The method of claim 1 , wherein each token in the sequence of tokens of the candidate transcription represents a word in the candidate transcription.
7. The method of claim 1 , wherein each token in the sequence of tokens of the candidate transcription represents a wordpiece in the candidate transcription.
8. The method of claim 1 , wherein each token in the sequence of tokens of the candidate transcription represents a n-gram, phoneme, or grapheme in the candidate transcription.
9. The method of claim 1 , wherein the first and second embedding tables are stored sparsely on memory hardware in communication with the data processing hardware.
10. The method of claim 1 , wherein determining the token embedding for the corresponding token comprises retrieving the token embedding from the first embedding table via a look-up without requiring access to any graphics processing units and/or tensor processing units.
11. The method of claim 1 , wherein the data processing hardware resides on the user device.
12. A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed by the data processing hardware cause the data processing hardware to perform operations comprising:
receiving audio data corresponding to an utterance spoken by a user and captured by a user device;
processing, using a speech recognizer, the audio data to determine a candidate transcription for the spoken utterance, the candidate transcription comprises a sequence of tokens;
for each corresponding token in the sequence of tokens subsequent to an initial token in the sequence of tokens:
determining, using a first embedding table, a token embedding for the corresponding token, the token embedding determined independent from each token in the sequence of tokens that precedes the corresponding token in the sequence of tokens;
determining, using a second embedding table, a n-gram token embedding for a previous sequence of n-gram tokens based on each token in the sequence of tokens that precedes the corresponding token in the sequence of tokens; and
concatenating the token embedding and the n-gram token embedding to generate a concatenated output for the corresponding token; and
rescoring, using an external language model, the candidate transcription for the spoken utterance by processing the concatenated output generated for each corresponding token in the sequence of tokens.
13. The system of claim 12 , wherein the external language model comprises a recurrent neural network language model.
14. The system of claim 12 , wherein the external language model integrates with the speech recognizer by Hybrid Autoregressive Transducer (HAT) factorizing the speech recognizer.
15. The system of claim 12 , wherein the speech recognizer comprises a conformer audio encoder and a recurrent neural network-transducer decoder.
16. The system of claim 12 , wherein the speech recognizer comprises a transformer audio encoder and a recurrent neural network-transducer decoder.
17. The system of claim 12 , wherein each token in the sequence of tokens of the candidate transcription represents a word in the candidate transcription.
18. The system of claim 12 , wherein each token in the sequence of tokens of the candidate transcription represents a wordpiece in the candidate transcription.
19. The system of claim 12 , wherein each token in the sequence of tokens of the candidate transcription represents a n-gram, phoneme, or grapheme in the candidate transcription.
20. The system of claim 12 , wherein the first and second embedding tables are stored sparsely on memory hardware in communication with the data processing hardware.
21. The system of claim 12 , wherein determining the token embedding for the corresponding token comprises retrieving the token embedding from the first embedding table via a look-up without requiring access to any graphics processing units and/or tensor processing units.
22. The system of claim 12 , wherein the data processing hardware resides on the user device.