System and method for speech recognition using deep recurrent neural networks
Deep recurrent neural networks applied to speech recognition. The deep recurrent neural networks (RNNs) are preferably implemented by stacked long short-term memory bidirectional RNNs. The RNNs are trained using end-to-end training with suitable regularisation.
1. A method performed by one or more computers and for training
a deep recurrent neural network (“deep RNN”) implemented on the one or more computers to perform speech recognition, the method comprising:
receiving training data for training the deep RNN, the training data comprising a plurality of sequences of audio observations and, for each sequence of audio observations, a corresponding sequence of text symbols that represents the sequence of audio observations comprising, and the deep RNN comprising:
a transcription neural network configured to generate a sequence of transcription hidden vectors that includes a respective transcription hidden vector for each audio observation in each sequence of audio observations;
a prediction neural network configured to, at position u in an output sequence generated from a particular sequence of audio observations, process a current output sequence that includes output labels at positions 1 through u−1 in the output sequence to generate a prediction hidden vector; and
an output neural network, configured to, at the position u in the output sequence generated from the particular sequence of audio observations, process the prediction hidden vector and a respective transcription hidden vector for an audio observation at a position t in the particular sequence of audio observations, to generate a probability distribution comprising a respective probability for each output label in a vocabulary of output labels; and
training the deep RNN to map the sequences of audio observations to the corresponding sequences of text symbols.
2. The method of claim 1 , wherein the transcription neural network is a bi-directional recurrent neural network.
3. The method of claim 1 , wherein the transcription neural network is a bi-directional long short-term memory (LSTM) neural network.
4. The method of claim 3 , wherein each transcription hidden vector is a concatenation of hidden vectors for the corresponding audio input generated by a set of uppermost bidirectional layers in the bi-directional LSTM neural network.
5. The method of claim 1 , wherein the prediction neural network is a uni-directional recurrent neural network.
6. The method of claim 1 , wherein the prediction neural network is a uni-directional long short-term memory (LSTM) neural network.
7. The method of claim 1 , wherein the vocabulary of output labels also includes a blank symbol that represents a non-output.
8. The method of claim 1 , wherein the transcription neural network is initialized from a CTC-trained neural network.
9. The method of claim 1 , wherein the prediction neural network is initialized from a next-step prediction network.
10. A system comprising one or more computers and one or more storage devices storing instructions for training a deep recurrent neural network (“deep RNN”) implemented on the one or more computers to perform speech recognition, the operations comprising:
receiving training data for training the deep RNN, the training data comprising a plurality of sequences of audio observations and, for each sequence of audio observations, a corresponding sequence of text symbols that represents the sequence of audio observations comprising, and the deep RNN comprising:
a transcription neural network configured to generate a sequence of transcription hidden vectors that includes a respective transcription hidden vector for each audio observation in each sequence of audio observations;
a prediction neural network configured to, at position u in an output sequence generated from a particular sequence of audio observations, process a current output sequence that includes output labels at positions 1 through u−1 in the output sequence to generate a prediction hidden vector; and
an output neural network, configured to, at the position u in the output sequence generated from the particular sequence of audio observations, process the prediction hidden vector and a respective transcription hidden vector for an audio observation at a position t in the particular sequence of audio observations, to generate a probability distribution comprising a respective probability for each output label in a vocabulary of output labels; and
training the deep RNN to map the sequences of audio observations to the corresponding sequences of text symbols.
11. The system of claim 10 , wherein the transcription neural network is a bi-directional recurrent neural network.
12. The system of claim 10 , wherein the transcription neural network is a bi-directional long short-term memory (LSTM) neural network.
13. The system of claim 12 , wherein each transcription hidden vector is a concatenation of hidden vectors for the corresponding audio input generated by a set of uppermost bidirectional layers in the bi-directional LSTM neural network.
14. The system of claim 10 , wherein the prediction neural network is a uni-directional recurrent neural network.
15. The system of claim 10 , wherein the prediction neural network is a uni-directional long short-term memory (LSTM) neural network.
16. The system of claim 10 , wherein the vocabulary of output labels also includes a blank symbol that represents a non-output.
17. The system of claim 10 , wherein the transcription neural network is initialized from a CTC-trained neural network.
18. The system of claim 10 , wherein the prediction neural network is initialized from a next-step prediction network.
19. One or more non-transitory computer-readable media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for training a deep recurrent neural network (“deep RNN”) implemented on the one or more computers to perform speech recognition, the operations comprising:
receiving training data for training the deep RNN, the training data comprising a plurality of sequences of audio observations and, for each sequence of audio observations, a corresponding sequence of text symbols that represents the sequence of audio observations comprising, and the deep RNN comprising:
a transcription neural network configured to generate a sequence of transcription hidden vectors that includes a respective transcription hidden vector for each audio observation in each sequence of audio observations;
a prediction neural network configured to, at position u in an output sequence generated from a particular sequence of audio observations, process a current output sequence that includes output labels at positions 1 through u−1 in the output sequence to generate a prediction hidden vector; and
an output neural network, configured to, at the position u in the output sequence generated from the particular sequence of audio observations, process the prediction hidden vector and a respective transcription hidden vector for an audio observation at a position t in the particular sequence of audio observations, to generate a probability distribution comprising a respective probability for each output label in a vocabulary of output labels; and
training the deep RNN to map the sequences of audio observations to the corresponding sequences of text symbols.
20. The one or more non-transitory computer-readable media of claim 19 , wherein the vocabulary of output labels also includes a blank symbol that represents a non-output.