Context-dependent state tying using a neural network

1. A method performed by one or more computers, the method comprising:

accessing, by the one or more computers, a first neural network that has been trained using speech data examples that are each assigned a context-dependent state by clustering context independent states based on activations at a hidden layer of a second neural network that was trained to provide outputs corresponding to context-independent states, the first neural network being configured to provide outputs corresponding to one or more context-dependent states;

receiving, by the one or more computers, an audio signal encoding a portion of an utterance;

providing, by the one or more computers, data corresponding to the audio signal to the first neural network that has been trained using the speech data examples that are each assigned a context-dependent state based on the activations at the hidden layer of the second neural network;

generating, by the one or more computers, data indicating a transcription for the utterance based on an output of the first neural network that was generated in response to the data corresponding to the audio signal; and

providing, by the one or more computers, the data indicating the transcription as output of an automated speech recognition service.

2. The method of claim 1 ,

wherein the second neural network is trained using vectors of acoustic data representing features of utterances.

3. The method of claim 2 , wherein the features include one or more phonemes.

4. The method of claim 1 , wherein the first neural network has been trained to provide outputs corresponding to multiple context-dependent states that are derived from a plurality of context-independent states using one or more decision trees.

5. The method of claim 4 , wherein at a given parent node of the one or more decision trees, a training frame corresponding to a context-independent state is assigned to one of a plurality of context-dependent child nodes.

6. The method of claim 4 , wherein the multiple context-dependent states are derived from the plurality of context-independent states using divisive, likelihood-based K-means clustering.

7. The method of claim 1 ,

wherein the method further comprises training the first neural network to provide outputs corresponding to context-dependent states after initializing the first neural network with parameter values from the second neural network trained to provide outputs corresponding to context-independent states.

8. The method of claim 1 , wherein the second neural network has been trained to indicate context-independent states corresponding to speech examples;

wherein the method further comprises:

using the second neural network to determine an activation vector for each of multiple speech examples, the activation vector for each speech sample indicating the activations produced at a hidden layer of the second neural network in response to inputting data about the speech sample to the second neural network;

assigning a context-dependent state to each of the multiple speech examples based on the activation vector for the speech sample; and

training the first neural network using the multiple speech examples and the context-dependent states assigned based on the activation vectors.

9. The method of claim 1 , further comprising deriving the first neural network from the second neural network by:

retaining hidden layers of the second neural network and parameters of the hidden layers of the second neural network;

replacing an output layer of the second neural network that provides outputs corresponding to context-independent states with an expanded output layer; and

training the expanded output layer to provide outputs corresponding to context-dependent states.

10. The method of claim 9 , wherein training the expanded output layer comprises adjusting parameters of the expanded output layer while maintaining the parameters of the hidden layers of the second neural network.

11. A system comprising

a speech recognition engine comprising one or more processors and a machine-readable storage device storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:

accessing, by the one or more computers, a first neural network that has been trained using speech examples that are each assigned a context-dependent state by clustering context independent states based on activations at a hidden layer of a second neural network that was trained to provide outputs corresponding to context-independent states, the first neural network being configured to provide outputs corresponding to one or more context-dependent states;

receiving, by the one or more computers, an audio signal encoding a portion of an utterance;

providing, by the one or more computers, data corresponding to the audio signal to the first neural network that has been trained using the speech examples that are each assigned a context-dependent state based on the activations at the hidden layer of the second neural network;

generating, by the one or more computers, data indicating a transcription for the utterance based on an output of the first neural network that was generated in response to the data corresponding to the audio signal; and

providing, by the one or more computers, the data indicating the transcription as output of an automated speech recognition service.

12. The system of claim 11 ,

wherein the second neural network is trained using vectors of acoustic data representing features of utterances.

13. The system of claim 12 , wherein the features include one or more phonemes.

14. The system of claim 11 , wherein the first neural network has been trained to provide outputs corresponding to multiple context-dependent states that are derived from a plurality of context-independent states using one or more decision trees.

15. The system of claim 14 , wherein at a given parent node of the one or more decision trees, a training frame corresponding to a context-independent state is assigned to one of a plurality of context-dependent child nodes.

16. The system of claim 14 , wherein the multiple context-dependent states are derived from the plurality of context-independent states using divisive, likelihood-based K-means clustering.

17. A non-transitory computer-readable storage device encoding one or more computer-readable instructions, which upon execution by one or more processors cause operations comprising:

accessing, by the one or more computers, a first neural network that has been trained using speech examples that are each assigned a context-dependent state by clustering context independent states based on activations at a hidden layer of a second neural network that was trained to provide outputs corresponding to context-independent states, the first neural network being configured to provide outputs corresponding to one or more context-dependent states;

receiving, by the one or more computers, an audio signal encoding a portion of an utterance;

providing, by the one or more computers, data corresponding to the audio signal to the first neural network that has been trained using the speech examples that are each assigned a context-dependent state based on the activations at the hidden layer of the second neural network;

generating, by the one or more computers, data indicating a transcription for the utterance based on an output of the first neural network that was generated in response to the data corresponding to the audio signal; and

providing, by the one or more computers, the data indicating the transcription as output of an automated speech recognition service.

18. The computer readable storage device of claim 17 ,

wherein the second neural network is trained using vectors of acoustic data representing features of utterances.

19. The computer readable storage device of claim 18 , wherein the features include one or more phonemes.