Generating representations of acoustic sequences
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating representation of acoustic sequences. One of the methods includes: receiving an acoustic sequence, the acoustic sequence comprising a respective acoustic feature representation at each of a plurality of time steps; processing the acoustic feature representation at an initial time step using an acoustic modeling neural network; for each subsequent time step of the plurality of time steps: receiving an output generated by the acoustic modeling neural network for a preceding time step, generating a modified input from the output generated by the acoustic modeling neural network for the preceding time step and the acoustic representation for the time step, and processing the modified input using the acoustic modeling neural network to generate an output for the time step; and generating a phoneme representation for the utterance from the outputs for each of the time steps.
1. A method comprising:
receiving, at an acoustic modeling system, an utterance comprising a respective acoustic feature representation at each time step in a time step sequence;
for each time step subsequent to an initial time step in the time step sequence:
generating, by the acoustic modeling system, a corresponding input to an acoustic modeling neural network by combining the respective acoustic feature representation at the time step and a preceding output from the acoustic modeling neural network for a preceding time step in the time step sequence; and
generating, by the acoustic modeling system, a corresponding probability distribution over possible phonemes for the time step by processing the input using the acoustic modeling neural network; and
generating, by the acoustic modeling system, a phoneme representation for the utterance based on the corresponding probability distribution over possible phonemes generated for each time step subsequent to the initial time step in the time step sequence.
2. The method of claim 1 , wherein the acoustic modeling neural network comprises a feed-forward neural network.
3. The method of claim 1 , wherein the acoustic modeling neural network comprises a recurrent neural network.
4. The method of claim 1 , wherein the acoustic modeling neural network comprises a long short-term memory (LSTM) neural network.
5. The method of claim 1 , wherein generating the phoneme representation for the utterance comprises:
for each time step subsequent to the initial time step in the time step sequence, selecting the phoneme associated with a highest probability from the corresponding probability distribution over possible phonemes for the time step; and
generating a sequence of the selected phonemes as the phoneme representation for the utterance.
6. The method of claim 1 , further comprising:
receiving, at the acoustic modeling system, a training acoustic sequence comprising a respective training acoustic feature representation at each of a plurality of training time steps, the training acoustic sequence associated with a known phoneme representation; and
training, by the acoustic modeling system, values of parameters of the acoustic modeling neural network based on the training acoustic sequence and the known phoneme representation.
7. The method of claim 1 , wherein the corresponding probability distribution over possible phonemes generated for each time step subsequent to the initial time step in the time step sequence represents a likelihood that each possible phoneme is a representation of the utterance at the time step.
8. The method of claim 1 , wherein the preceding output from the acoustic modeling neural network for the preceding time step in the time step sequence comprises a probability distribution over possible phonemes for the preceding time step.
9. The method of claim 8 , wherein generating the corresponding input to the acoustic modeling neural network for each time step subsequent to the initial time step in the time step sequence comprises appending the probability distribution over possible phonemes for the preceding time step to the respective acoustic feature representation at the time step.
10. The method of claim 8 , wherein generating the corresponding input to the acoustic modeling neural network for each time step subsequent to the initial time step in the time step sequence comprises appending a phoneme associated with a highest probability from the corresponding probability distribution over possible phonemes for the preceding time step.
11. The method of claim 1 , further comprising, for the initial time step in the time step sequence, generating, by the acoustic modeling system, a corresponding output by processing a predetermined input using the acoustic modeling neural network.
12. The method of claim 11 , wherein the predetermined input comprises a vector of predetermined values.
13. The method of claim 11 , wherein the predetermined input comprises a vector of zeroes.
14. The method of claim 1 , wherein the respective acoustic feature representation at each time step in the time step sequence comprises a respective acoustic feature vector.
15. The method of claim 1 , further comprising providing the generated phoneme representation for the utterance to a speech recognition system.
16. An acoustic modeling system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware and storing instructions that when executed by the data processing hardware cause the data processing hardware to perform operations comprising:
receiving an utterance comprising a respective acoustic feature representation at each time step in a time step sequence;
for each time step subsequent to an initial time step in the time step sequence:
generating a corresponding input to an acoustic modeling neural network by combining the respective acoustic feature representation at the time step and a preceding output from the acoustic modeling neural network for a preceding time step in the time step sequence; and
generating a corresponding probability distribution over possible phonemes for the time step by processing the input using the acoustic modeling neural network; and
generating a phoneme representation for the utterance based on the corresponding probability distribution over possible phonemes generated for each time step subsequent to the initial time step in the time step sequence.
17. The acoustic modeling system of claim 16 , wherein the acoustic modeling neural network comprises a feed-forward neural network.
18. The acoustic modeling system of claim 16 , wherein the acoustic modeling neural network comprises a recurrent neural network.
19. The acoustic modeling system of claim 16 , wherein the acoustic modeling neural network comprises a long short-term memory (LSTM) neural network.
20. The acoustic modeling system of claim 16 , wherein generating the phoneme representation for the utterance comprises:
for each time step subsequent to the initial time step in the time step sequence, selecting the phoneme associated with a highest probability from the corresponding probability distribution over possible phonemes for the time step; and
generating a sequence of the selected phonemes as the phoneme representation for the utterance.
21. The acoustic modeling system of claim 16 , wherein the operations further comprise:
receiving a training acoustic sequence comprising a respective training acoustic feature representation at each of a plurality of training time steps, the training acoustic sequence associated with a known phoneme representation; and
training values of parameters of the acoustic modeling neural network based on the training acoustic sequence and the known phoneme representation.
22. The acoustic modeling system of claim 16 , wherein the corresponding probability distribution over possible phonemes generated for each time step subsequent to the initial time step in the time step sequence represents a likelihood that each possible phoneme is a representation of the utterance at the time step.
23. The acoustic modeling system of claim 16 , wherein the preceding output from the acoustic modeling neural network for the preceding time step in the time step sequence comprises a probability distribution over possible phonemes for the preceding time step.
24. The acoustic modeling system of claim 23 , wherein generating the corresponding input to the acoustic modeling neural network for each time step subsequent to the initial time step in the time step sequence comprises appending the probability distribution over possible phonemes for the preceding time step to the respective acoustic feature representation at the time step.
25. The acoustic modeling system of claim 23 , wherein generating the corresponding input to the acoustic modeling neural network for each time step subsequent to the initial time step in the time step sequence comprises appending a phoneme associated with a highest probability from the corresponding probability distribution over possible phonemes for the preceding time step.
26. The acoustic modeling system of claim 16 , wherein the operations further comprise, for the initial time step in the time step sequence, generating, a corresponding output by processing a predetermined input using the acoustic modeling neural network.
27. The acoustic modeling system of claim 26 , wherein the predetermined input comprises a vector of predetermined values.
28. The acoustic modeling system of claim 26 , wherein the predetermined input comprises a vector of zeroes.
29. The acoustic modeling system of claim 16 , wherein the respective acoustic feature representation at each time step in the time step sequence comprises a respective acoustic feature vector.
30. The acoustic modeling system of claim 16 , wherein the operations further comprise providing the generated phoneme representation for the utterance to a speech recognition system.