Methods and apparatus for training a transformation component

1. A method of adapting a trained acoustic model, which was trained using first training data including speech data obtained over a near-field channel type, to process speech data obtained over a far-field channel type, the trained acoustic model comprising first parameters having respective first values established during training of the trained acoustic model using the first training data, wherein the trained acoustic model comprises a multi-layer neural network, the method comprising:

using at least one computer processor to perform:

obtaining second training data comprising speech data obtained over the far-field channel type without obtaining corresponding stereo data over the near-field channel type; and

adapting the trained acoustic model to process speech data obtained over the far-field channel type by:

augmenting the trained acoustic model with a transformation component configured to receive input derived from speech data obtained over the far-field channel type, apply a transformation to the input to obtain transformed input, and provide the transformed input as input to the trained acoustic model, the transformation component comprising second parameters and at least one network layer, wherein the transformation component is configured to linearly transform the input to obtain the transformed input, wherein the augmenting comprises coupling outputs of the at least one network layer of the transformation component to inputs of a first layer of the multi-layer neural network; and

training the transformation component by using only the second training data to determine respective second values for the second parameters, wherein training the transformation component by using only the second training data to determine the respective second values for the second parameters comprises:

comparing sequence data output by the trained acoustic model in response to the transformed input with an expected sequence data output; and

adjusting the second values for the second parameters based on the comparison of the sequence data output and the expected sequence data output.

2. The method of claim 1 , wherein the trained acoustic model comprises a deep neural network and the at least one network layer of the transformation component comprises a linear input network, and wherein the augmenting comprises coupling outputs of the linear input network to inputs of an input layer of the deep neural network.

3. The method of claim 1 , wherein the second training data is obtained independently of the first training data.

4. At least one non-transitory computer readable storage medium storing instructions that, when executed by at least one hardware processor, perform a method of adapting a trained acoustic model, which was trained using first training data including speech data obtained over a near-field channel type, to process speech data obtained over a far-field channel type, the trained acoustic model comprising first parameters having respective first values established during training of the trained acoustic model using the first training data, wherein the trained acoustic model comprises a multi-layer neural network, the method comprising:

obtaining second training data comprising speech data obtained over the far-field channel type without obtaining corresponding stereo data over the near-field channel type; and

adapting the trained acoustic model to process speech data obtained over the far-field channel type by:

augmenting the trained acoustic model with a transformation component configured to receive input derived from speech data obtained over the far-field channel type, apply a transformation to the input to obtain transformed input, and provide the transformed input as input to the trained acoustic model, the transformation component comprising second parameters and at least one network layer, wherein the transformation component is configured to linearly transform the input to obtain the transformed input, wherein the augmenting comprises coupling outputs of the at least one network layer of the transformation component to inputs of a first layer of the multi-layer neural network; and

training the transformation component by using only the second training data to

determine respective second values for the second parameters, wherein training the transformation component by using only the second training data to determine the respective second values for the second parameters comprises:

comparing sequence data output by the trained acoustic model in response to the transformed input with an expected sequence data output; and

adjusting the second values for the second parameters based on the comparison of the sequence data output and the expected sequence data output.

5. The at least one non-transitory computer readable medium of claim 4 , wherein the trained acoustic model comprises a deep neural network and the at least one network layer of the transformation component comprises a linear input network, and wherein the augmenting comprises coupling outputs of the linear input network to inputs of an input layer of the deep neural network.

6. The at least one non-transitory computer readable medium of claim 4 , wherein the second training data is obtained independently of the first training data.

7. A system configured adapt a trained acoustic model, which was trained using first training data including speech data obtained over a near-field channel type, to process speech data obtained over a far-field channel type, the trained acoustic model comprising first parameters having respective first values established during training of the trained acoustic model using the first training data, wherein the trained acoustic model comprises a multi-layer neural network, the system comprising:

at least one hardware processor configured to:

obtaining second training data comprising speech data obtained over the far-field channel type without obtaining corresponding stereo data over the near-field channel type; and

adapting the trained acoustic model to process speech data obtained over the far-field channel type by:

augmenting the trained acoustic model with a transformation component configured to receive input derived from speech data obtained over the far-field channel type, apply a transformation to the input to obtain transformed input, and provide the transformed input as input to the trained acoustic model, the transformation component comprising second parameters and at least one network layer, wherein the transformation component is configured to linearly transform the input to obtain the transformed input, wherein the augmenting comprises coupling outputs of the at least one network layer of the transformation component to inputs of a first layer of the multi-layer neural network; and

training the transformation component by using only the second training data to determine respective second values for the second parameters, wherein training the transformation component by using only the second training data to determine the respective second values for the second parameters comprises:

comparing sequence data output by the trained acoustic model in response to the transformed input with an expected sequence data output; and

adjusting the second values for the second parameters based on the comparison of the sequence data output and the expected sequence data output.

8. The system of claim 7 , wherein the second training data is obtained independently of the first training data.

9. The method of claim 1 , wherein the trained acoustic model retains the first parameters having the respective first values throughout training of the transformation component.

10. The at least one non-transitory computer-readable storage medium of claim 4 , wherein the trained acoustic model retains the first parameters having the respective first values throughout training of the transformation component.

11. The system of claim 7 , wherein the trained acoustic model retains the first parameters having the respective first values throughout training of the transformation component.

12. The method of claim 1 ,

wherein the second training data further comprises expected sequence data corresponding to the speech data obtained over the far-field channel type, the expected sequence data comprising sub-word speech units corresponding to the acoustic speech data in the second training data, and

wherein training the transformation component comprises determining the respective second values for the second parameters at least in part by:

transforming at least some of the speech data obtained over the far-field channel type to obtain transformed speech data;

providing the transformed speech data as input to the trained acoustic model to obtain output sequence data; and

comparing output sequence data to the expected sequence data.

13. The at least one non-transitory computer readable storage medium of claim 4 ,

wherein the second training data further comprises expected sequence data corresponding to the speech data obtained over the far-field channel type, the expected sequence data comprising sub-word speech units corresponding to the acoustic speech data in the second training data, and

wherein training the transformation component comprises determining the respective second values for the second parameters at least in part by:

transforming at least some of the speech data obtained over the far-field channel type to obtain transformed speech data;

providing the transformed speech data as input to the trained acoustic model to obtain output sequence data; and

comparing output sequence data to the expected sequence data.

14. The system of claim 7 ,

wherein the second training data further comprises expected sequence data corresponding to the speech data obtained over the far-field channel type, the expected sequence data comprising sub-word speech units corresponding to the acoustic speech data in the second training data, and

wherein training the transformation component comprises determining the respective second values for the second parameters at least in part by:

transforming at least some of the speech data obtained over the far-field channel type to obtain transformed speech data;

providing the transformed speech data as input to the trained acoustic model to obtain output sequence data; and

comparing output sequence data to the expected sequence data.