Emitting word timings with end-to-end models
A method includes receiving a training example that includes audio data representing a spoken utterance and a ground truth transcription. For each word in the spoken utterance, the method also includes inserting a placeholder symbol before the respective word identifying a respective ground truth alignment for a beginning and an end of the respective word, determining a beginning word piece and an ending word piece, and generating a first constrained alignment for the beginning word piece and a second constrained alignment for the ending word piece. The first constrained alignment is aligned with the ground truth alignment for the beginning of the respective word and the second constrained alignment is aligned with the ground truth alignment for the ending of the respective word. The method also includes constraining an attention head of a second pass decoder by applying the first and second constrained alignments.
1. A computer-implemented method when executed on data processing hardware causes the data processing hardware to perform operations comprising:
receiving audio data characterizing a spoken utterance;
encoding the audio data into a sequence of audio encodings;
processing, using a second pass decoder of a two-pass neural network model, the sequence of audio encodings to generate a sequence of word pieces that form a transcription of one or more words of the spoken utterance;
for each word in the transcription of the one or more words of the spoken utterance, generating a first constrained alignment for the beginning word piece of the respective word and a second constrained alignment for the ending word piece of the respective word;
constraining an attention head of the second pass decoder of the two-pass neural network model based on the first constrained alignment and the second constrained alignment generated for each word in the transcription; and
determining, using the constrained attention head of the second pass decoder, actual word timings for each word in the transcription of the one or more words of the spoken utterance.
2. The computer-implemented method of claim 1 , wherein the operations further comprise processing, using a first pass decoder of the two-pass neural network model, the sequence of audio encodings to generate a plurality of streaming speech recognition hypotheses for the spoken utterance.
3. The computer-implemented method of claim 2 , wherein processing the sequence of audio encodings to generate the sequence of word pieces using the second pass decoder of the two-pass neural network model further comprises processing the sequence of audio encodings and the top-K streaming speech recognition hypotheses among the plurality of streaming speech recognition hypotheses to generate the sequence of word pieces that form the transcription of one or more words of the spoken utterance.
4. The computer-implemented method of claim 2 , wherein, when processing the sequence of audio encodings to generate the sequence of word pieces, the second pass decoder of the two-pass neural network ignores the plurality of streaming speech recognition hypotheses.
5. The computer-implemented method of claim 2 , wherein the first pass decoder comprises a prediction network and a joint network.
6. The computer-implemented method of claim 2 , wherein:
the first pass decoder of the two-pass neural network model comprises a recurrent neural network-transducer (RNN-T) decoder; and
the second pass decoder of the two-pass neural network model comprises a Listen, Attend, and Spell (LAS) decoder.
7. The computer-implemented method of claim 1 , wherein the data processing hardware resides on a user device that captured the spoken utterance in streaming audio.
8. The computer-implemented method of claim 1 , wherein the operations further comprise displaying, on a screen in communication with the data processing hardware, the transcription of one or more words of the spoken utterance, the transcription annotated with the actual word timings determined for each word in the transcription.
9. The computer-implemented method of claim 1 , wherein the second pass decoder comprises a plurality of attention heads.
10. The computer-implemented method of claim 1 , wherein constraining the attention head comprises constraining an attention probability derived from the attention head of the second pass decoder.
11. 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 characterizing a spoken utterance;
encoding the audio data into a sequence of audio encodings;
processing, using a second pass decoder of a two-pass neural network model, the sequence of audio encodings to generate a sequence of word pieces that form a transcription of one or more words of the spoken utterance;
for each word in the transcription of the one or more words of the spoken utterance, generating a first constrained alignment for the beginning word piece of the respective word and a second constrained alignment for the ending word piece of the respective word;
constraining an attention head of the second pass decoder of the two-pass neural network model based on the first constrained alignment and the second constrained alignment generated for each word in the transcription; and
determining, using the constrained attention head of the second pass decoder, actual word timings for each word in the transcription of the one or more words of the spoken utterance.
12. The system of claim 11 , wherein the operations further comprise processing, using a first pass decoder of the two-pass neural network model, the sequence of audio encodings to generate a plurality of streaming speech recognition hypotheses for the spoken utterance.
13. The system of claim 12 , wherein processing the sequence of audio encodings to generate the sequence of word pieces using the second pass decoder of the two-pass neural network model further comprises processing the sequence of audio encodings and the top-K streaming speech recognition hypotheses among the plurality of streaming speech recognition hypotheses to generate the sequence of word pieces that form the transcription of one or more words of the spoken utterance.
14. The system of claim 12 , wherein, when processing the sequence of audio encodings to generate the sequence of word pieces, the second pass decoder of the two-pass neural network ignores the plurality of streaming speech recognition hypotheses.
15. The system of claim 12 , wherein the first pass decoder comprises a prediction network and a joint network.
16. The system of claim 12 , wherein:
the first pass decoder of the two-pass neural network model comprises a recurrent neural network-transducer (RNN-T) decoder; and
the second pass decoder of the two-pass neural network model comprises a Listen, Attend, and Spell (LAS) decoder.
17. The system of claim 11 , wherein the data processing hardware resides on a user device that captured the spoken utterance in streaming audio.
18. The system of claim 11 , wherein the operations further comprise displaying, on a screen in communication with the data processing hardware, the transcription of one or more words of the spoken utterance, the transcription annotated with the actual word timings determined for each word in the transcription.
19. The system of claim 11 , wherein the second pass decoder comprises a plurality of attention heads.
20. The system of claim 11 , wherein constraining the attention head comprises constraining an attention probability derived from the attention head of the second pass decoder.