Speech recognition using convolutional neural networks
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing speech recognition by generating a neural network output from an audio data input sequence, where the neural network output characterizes words spoken in the audio data input sequence. One of the methods includes, for each of the audio data inputs, providing a current audio data input sequence that comprises the audio data input and the audio data inputs preceding the audio data input in the audio data input sequence to a convolutional subnetwork comprising a plurality of dilated convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of audio data inputs: receive the current audio data input sequence for the audio data input, and process the current audio data input sequence to generate an alternative representation for the audio data input.
1. A neural network system implemented by one or more computers,
wherein the neural network system is configured to perform speech recognition by processing an audio data input sequence comprising a plurality of audio data inputs to generate a neural network output which characterizes words spoken in the audio data input sequence, and
wherein the neural network system comprises:
a convolutional subnetwork comprising a plurality of dilated convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of audio data inputs:
receive a current audio data input sequence comprising the audio data input and the audio data inputs that precede the audio data input in the audio data input sequence, and
process the current audio data input sequence to generate an alternative representation for the audio data input, wherein the dilated convolutional neural network layers are causal convolutional neural network layers and the alternative representation for the audio data input does not depend on any audio data inputs that follow the audio data input in the audio data input sequence; and
an output subnetwork, wherein the output subnetwork is configured to receive the alternative representations of the audio data inputs and to process the alternative representations to generate the neural network output characterizing words spoken in the audio data input sequence.
2. The neural network system of claim 1 ,
wherein the neural network output comprises a plurality of sets of scores,
wherein each set of scores includes a respective score for each of a plurality of possible outputs, and
wherein the neural network system further comprises:
a subsystem configured to select outputs in accordance with the sets of scores to generate an output sequence comprising a plurality of outputs.
3. The neural network system of claim 2 , wherein the output sequence is a sequence of graphemes that is a grapheme representation of words spoken in the audio data input sequence.
4. The neural network system of claim 2 , wherein the output sequence is a sequence of phonemes that is a phoneme representation of words spoken in the audio data input sequence.
5. The neural network system of claim 2 , wherein the output sequence is a sequence of words in a first natural language that represents worked spoken in the audio data input sequence.
6. The neural network system of claim 2 , wherein the output subnetwork comprises:
a mean pooling layer configured to:
receive a subset of the alternative representations, and
perform mean pooling on the subset of alternative representations to generate a mean pooled representation for the subset; and
one or more additional neural network layers configured to, for each subset of the alternative representations, receive the mean pooled representation for the subset and to generate a set of scores for a position in the output sequence corresponding to the subset.
7. The neural network system of claim 1 , wherein the dilated convolutional neural network layers include multiple blocks of dilated convolutional neural network layers, wherein each block comprises multiple dilated convolutional neural network layers with increasing dilation.
8. The neural network system of claim 1 , wherein one or more of the dilated convolutional neural network layers have gated activation units.
9. The neural network system of claim 1 , wherein the convolutional subnetwork comprises residual connections.
10. The neural network system of claim 1 , wherein the convolutional subnetwork comprises skip connections.
11. A method of performing speech recognition by generating a neural network output from an audio data input sequence that comprises a plurality of audio data inputs, wherein the neural network output characterizes words spoken in the audio data input sequence, the method comprising:
for each of the audio inputs, providing a current audio data input sequence that comprises the audio data input and the audio data inputs preceding the audio data input in the audio data input sequence to a convolutional subnetwork comprising a plurality of dilated convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of audio data inputs:
receive the current audio data input sequence for the audio data input, and
process the current audio data input sequence to generate an alternative representation for the audio data input, wherein the dilated convolutional neural network layers are causal convolutional neural network layers and the alternative representation for the audio data input does not depend on any audio data inputs that follow the audio data input in the audio data input sequence; and
providing the alternative representations to an output subnetwork, wherein the output subnetwork is configured to receive the alternative representations and to process the alternative representations to generate the neural network output characterizing words spoken in the audio data input sequence.
12. The method of claim 11 , wherein the neural network output comprises a plurality of sets of scores,
wherein each set of scores includes a respective score for each of a plurality of possible outputs, and
wherein the method further comprises:
selecting outputs in accordance with the sets of scores to generate an output sequence comprising a plurality of outputs.
13. The method of claim 12 , wherein the output sequence is a sequence of graphemes that is a grapheme representation of the audio data input sequence.
14. The method of claim 12 , wherein the output sequence is a sequence of phonemes that is a phoneme representation of the audio data input sequence.
15. The method of claim 12 , wherein the output sequence is a sequence of words in a first natural language that represents the audio data input sequence.
16. The method of claim 12 , wherein the output subnetwork comprises:
a mean pooling layer configured to:
receive a subset of the alternative representations, and
perform mean pooling on the subset of alternative representations to generate a mean pooled representation for the subset; and
one or more additional neural network layers configured to, for each subset of the alternative representations, receive the mean pooled representation for the subset and to generate a subset output that defines an output in the output sequence at a position corresponding to the sub set.
17. The method of claim 11 , wherein the dilated convolutional neural network layers include multiple blocks of dilated convolutional neural network layers, wherein each block comprises multiple dilated convolutional neural network layers with increasing dilation.
18. The method of claim 11 , wherein one or more of the dilated convolutional neural network layers have gated activation units.
19. The method of claim 11 , wherein the convolutional subnetwork comprises residual connections.
20. The method of claim 11 , wherein the convolutional subnetwork comprises skip connections.
21. One or more non-transitory computer storage media encoded with instructions that when executed by one or more computers cause the one or more computers to perform speech recognition operations by generating a neural network output from an audio data input sequence that comprises a plurality of audio data inputs, wherein the neural network output characterizes words spoken in the audio data input sequence, the speech recognition operations comprising:
for each of the audio inputs, providing a current audio data input sequence that comprises the audio data input and the audio data inputs preceding the audio data input in the audio data input sequence to a convolutional subnetwork comprising a plurality of dilated convolutional neural network layers, wherein the convolutional subnetwork is configured to, for each of the plurality of audio data inputs:
receive the current audio data input sequence for the audio data input, and
process the current audio data input sequence to generate an alternative representation for the audio data input, wherein the dilated convolutional neural network layers are causal convolutional neural network layers and the alternative representation for the audio data input does not depend on any audio data inputs that follow the audio data input in the audio data input sequence; and
providing the alternative representations to an output subnetwork, wherein the output subnetwork is configured to receive the alternative representations and to process the alternative representations to generate the neural network output characterizing words spoken in the audio data input sequence.