Processing audio waveforms
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing audio waveforms. In some implementations, a time-frequency feature representation is generated based on audio data. The time-frequency feature representation is input to an acoustic model comprising a trained artificial neural network. The trained artificial neural network comprising a frequency convolution layer, a memory layer, and one or more hidden layers. An output that is based on output of the trained artificial neural network is received. A transcription is provided, where the transcription is determined based on the output of the acoustic model.
1. A method comprising:
obtaining audio waveform samples from audio data describing an utterance;
providing the audio waveform samples as input to a trained neural network that has been trained to classify speech,
wherein the trained neural network comprises a feature extraction portion that comprises at least one layer of the trained neural network,
wherein the trained neural network comprises a classification portion comprising at least one layer of the trained neural network, and
wherein the feature extraction portion is configured to (i) generate a plurality of values based on the audio waveform samples from the audio data using parameters determined through training of the trained neural network, and (ii) generate a feature vector for output to the classification portion that has a reduced dimensionality with respect to the plurality of values;
receiving output data that the trained neural network generates in response to receiving the audio waveform samples, wherein the output data indicates likelihoods that different speech units have occurred; and
using the output data from the trained neural network to identify one or more words included in the utterance.
2. The method of claim 1 , wherein the output data includes a value that indicates a likelihood that a portion of the utterance represents a particular context-dependent state.
3. The method of claim 1 , wherein obtaining the audio waveform samples comprises identifying audio waveform samples corresponding to a frame of audio data; and
wherein the feature extraction portion is configured to generate a frame-level feature vector corresponding to the audio waveform samples corresponding to the frame of audio data.
4. The method of claim 1 , wherein the classification portion is configured to receive feature values determined using the feature extraction portion and generate the output data.
5. The method of claim 1 , wherein the classification portion comprises multiple hidden layers of the trained neural network.
6. The method of claim 1 , wherein the classification portion is configured to provide, as the output data, a likelihood corresponding to each of multiple different output labels that each represent a different speech unit.
7. The method of claim 1 , wherein the feature extraction portion and the classification portion each include at least one layer of the trained neural network, wherein parameters of the feature extraction portion and the classification portion are trained jointly during training of the trained neural network.
8. The method of claim 7 , wherein the trained neural network has been trained using a cross-entropy criterion, using asynchronous stochastic gradient descent (ASGD) optimization.
9. The method of claim 1 , wherein the speech units comprise a predetermined set of phonetic units.
10. The method of claim 9 , wherein the predetermined set of phonetic units comprises context-dependent phonetic units.
11. The method of claim 9 , wherein the predetermined set of phonetic units comprise hidden markov model (HMM) states.
12. The method of claim 11 , wherein the HMM states correspond to phonemes or portions of phonemes.
13. The method of claim 1 , comprising providing an output indicating an occurrence in the utterance of the identified one or more words.
14. The method of claim 1 , wherein the feature extraction portion uses at least one finite impulse response filter to generate a feature representation from the audio waveform samples.
15. The method of claim 14 , wherein the at least one finite impulse response filter has parameters trained jointly with training of other parameters of the trained neural network.
16. The method of claim 14 , wherein the feature extraction portion is configured to use multiple finite impulse response filters to generate feature representations used to generate values provided as input to one or more layers of the trained neural network.
17. The method of claim 1 , wherein the feature extraction portion comprises a convolutional neural network layer.
18. The method of claim 17 , wherein the convolutional neural network layer is configured to perform a time-domain convolution.
19. A system comprising:
one or more computers; and
one or more computer-readable media storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
obtaining audio waveform samples from audio data describing an utterance;
providing the audio waveform samples as input to a trained neural network that has been trained to classify speech,
wherein the trained neural network comprises a feature extraction portion that comprises at least one layer of the trained neural network,
wherein the trained neural network comprises a classification portion comprising at least one layer of the trained neural network, and
wherein the feature extraction portion is configured to (i) generate a plurality of values based on the audio waveform samples from the audio data using parameters determined through training of the trained neural network, and (ii) generate a feature vector for output to the classification portion that has a reduced dimensionality with respect to the plurality of values;
receiving output data that the trained neural network generates in response to receiving the audio waveform samples, wherein the output data indicates likelihoods that different speech units have occurred; and
using the output data from the trained neural network to identify one or more words included in the utterance.
20. 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 comprising:
obtaining audio waveform samples from audio data describing an utterance;
providing the audio waveform samples as input to a trained neural network that has been trained to classify speech,
wherein the trained neural network comprises a feature extraction portion that comprises at least one layer of the trained neural network,
wherein the trained neural network comprises a classification portion comprising at least one layer of the trained neural network, and
wherein the feature extraction portion is configured to (i) generate a plurality of values based on the audio waveform samples from the audio data using parameters determined through training of the trained neural network, and (ii) generate a feature vector for output to the classification portion that has a reduced dimensionality with respect to the plurality of values;
receiving output data that the trained neural network generates in response to receiving the audio waveform samples, wherein the output data indicates likelihoods that different speech units have occurred; and
using the output data from the trained neural network to identify one or more words included in the utterance.