Deliberation model-based two-pass end-to-end speech recognition
A method of performing speech recognition using a two-pass deliberation architecture includes receiving a first-pass hypothesis and an encoded acoustic frame and encoding the first-pass hypothesis at a hypothesis encoder. The first-pass hypothesis is generated by a recurrent neural network (RNN) decoder model for the encoded acoustic frame. The method also includes generating, using a first attention mechanism attending to the encoded acoustic frame, a first context vector, and generating, using a second attention mechanism attending to the encoded first-pass hypothesis, a second context vector. The method also includes decoding the first context vector and the second context vector at a context vector decoder to form a second-pass hypothesis.
1. A computer-implemented method when executed on data processing hardware causes the data processing hardware to perform operations comprising:
receiving a first-pass hypothesis and an encoded acoustic frame, the first-pass hypothesis generated by a recurrent neural network (RNN) decoder model for the encoded acoustic frame;
encoding the first-pass hypothesis at a hypothesis encoder;
generating, using an acoustic attention mechanism attending to the encoded acoustic frame, a first context vector;
generating, using a hypothesis attention mechanism attending to the encoded first-pass hypothesis, a second context vector; and
decoding the first context vector and the second context vector at a context vector decoder to form a second-pass hypothesis.
2. The computer-implemented method of claim 1 , wherein decoding the first context vector and the second context vector comprises decoding a concatenation of the first context vector and the second context vector.
3. The computer-implemented method of claim 1 , wherein encoding the first-pass hypothesis comprises bi-directionally encoding the first-pass hypothesis at the hypothesis encoder to generate contextual information from the first-pass hypothesis.
4. The computer-implemented method of claim 3 , wherein the hypothesis encoder comprises a long short term memory (LSTM) network.
5. The computer-implemented method of claim 1 , wherein the operations further comprise:
encoding the acoustic frame at a shared encoder; and
generating the first-pass hypothesis at the RNN decoder model based on the encoded acoustic frame communicated from the shared encoder.
6. The computer-implemented method of claim 5 , wherein the operations further comprise generating an acoustic embedding at a unidirectional audio encoder based on the encoded acoustic frame communicated from the shared encoder.
7. The computer-implemented method of claim 6 , wherein the unidirectional audio encoder comprises a long short term memory (LSTM) network.
8. The computer-implemented method of claim 7 , wherein the LSTM network comprises at least two layers.
9. The computer-implemented method of claim 1 , wherein the operations further comprise:
training the RNN decoder model; and
training a deliberation decoder while parameters of the trained RNN decoder model remain fixed, the deliberation decoder comprising the hypothesis encoder, the acoustic attention mechanism, the hypothesis attention mechanism, and the context vector decoder.
10. The computer-implemented method of claim 9 , wherein the operations further comprise minimizing a word error rate during training of the RNN decoder model and the deliberation decoder model.
11. The computer-implemented method of claim 1 , wherein the operations further comprise jointly training the RNN decoder model and a deliberation decoder, the deliberation decoder comprising the hypothesis encoder, the acoustic attention mechanism, the hypothesis attention mechanism, and the context vector decoder.
12. The computer-implemented method of claim 11 , wherein the operations further comprise minimizing a word error rate during the joint training of the RNN decoder model and the deliberation decoder model.
13. The computer-implemented method of claim 1 , wherein the data processing hardware resides on a user device.
14. 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 a first-pass hypothesis and an encoded acoustic frame, the first-pass hypothesis generated by a recurrent neural network (RNN) decoder model for the encoded acoustic frame;
encoding the first-pass hypothesis at a hypothesis encoder;
generating, using an acoustic attention mechanism attending to the encoded acoustic frame, a first context vector;
generating, using a hypothesis attention mechanism attending to the encoded first-pass hypothesis, a second context vector; and
decoding the first context vector and the second context vector at a context vector decoder to form a second-pass hypothesis.
15. The system of claim 14 , wherein decoding the first context vector and the second context vector comprises decoding a concatenation of the first context vector and the second context vector.
16. The system of claim 14 , wherein encoding the first-pass hypothesis comprises bi-directionally encoding the first-pass hypothesis at the hypothesis encoder to generate contextual information from the first-pass hypothesis.
17. The system of claim 16 , wherein the hypothesis encoder comprises a long short term memory (LSTM) network.
18. The system of claim 14 , wherein the operations further comprise:
encoding the acoustic frame at a shared encoder; and
generating the first-pass hypothesis at the RNN decoder model based on the encoded acoustic frame communicated from the shared encoder.
19. The system of claim 18 , wherein the operations further comprise generating an acoustic embedding at a unidirectional audio encoder based on the encoded acoustic frame communicated from the shared encoder.
20. The system of claim 19 , wherein the unidirectional audio encoder comprises a long short term memory (LSTM) network.
21. The system of claim 20 , wherein the LSTM network comprises at least two layers.
22. The system of claim 14 , wherein the operations further comprise:
training the RNN decoder model; and
training a deliberation decoder while parameters of the trained RNN decoder model remain fixed, the deliberation decoder comprising the hypothesis encoder, the acoustic attention mechanism, the hypothesis attention mechanism, and the context vector decoder.
23. The system of claim 22 , wherein the operations further comprise minimizing a word error rate during training of the RNN decoder model and the deliberation decoder model.
24. The system of claim 14 , wherein the operations further comprise jointly training the RNN decoder model and a deliberation decoder, the deliberation decoder comprising the hypothesis encoder, the acoustic attention mechanism, the hypothesis attention mechanism, and the context vector decoder.
25. The system of claim 24 , wherein the operations further comprise minimizing a word error rate during the joint training of the RNN decoder model and the deliberation decoder model.
26. The system of claim 14 , wherein the data processing hardware resides on a user device.