Conservatively adapting a deep neural network in a recognition system
AssigneeMicrosoft Technology Licensing, LLC
InventorsDong Yu, Kaisheng Yao, Hang Su, Gang Li, Frank Seide
Life of this patent
Estimating the term…
Abstract
Various technologies described herein pertain to conservatively adapting a deep neural network (DNN) in a recognition system for a particular user or context. A DNN is employed to output a probability distribution over models of context-dependent units responsive to receipt of captured user input. The DNN is adapted for a particular user based upon the captured user input, wherein the adaption is undertaken conservatively such that a deviation between outputs of the adapted DNN and the unadapted DNN is constrained.
Claims20
A computing device comprising a recognition system, the recognition system comprising:
- a deep neural network (DNN) trained using training data corresponding to a plurality of users, the DNN parameterized by a plurality of parameters and comprising multiple layers, the multiple layers comprising:
- an input layer that receives features of a captured observation;
- a plurality of hidden layers, each hidden layer comprising a respective plurality of nodes, each node in a hidden layer being configured to perform a nonlinear transformation on output of at least one node from an adjacent layer in the DNN;
- and
- an output layer that comprises models of context-dependent units, the output layer receiving outputs of nodes from an uppermost hidden layer in the DNN, wherein output of the DNN comprises a probability distribution over the models of the context-dependent units;
- and
- an adapter component that conservatively adapts at least one parameter of the plurality of parameters to form an adapted DNN, wherein the adaptor component conservatively adapts the at least one parameter of the plurality of parameters such that deviation between outputs of the DNN and the adapted DNN is constrained.
A method executed by a computing device comprising a speech recognition system, the speech recognition system including a deep neural network (DNN) trained using speech data from a plurality of users, the DNN comprising a plurality of learned parameters, the method comprising:
- receiving a spoken utterance of a particular user;
- responsive to receiving the spoken utterance, selecting a value of a regularization weight based at least in part upon a length of the spoken utterance in time;
- and
- responsive to selecting the value of the regularization weight, adapting at least one parameter of the plurality of learned parameters in the DNN based at least in part upon the regularization weight and the spoken utterance to form an adapted DNN, the adapted DNN being customized for the particular user.
A computer-readable storage medium comprising instructions that, when executed by a processor, cause the processor to perform acts comprising:
- receiving a spoken utterance from a particular user;
- responsive to receiving the spoken utterance, accessing a speaker-independent deep neural network (SI-DNN), wherein the SI-DNN comprises at least three hidden layers and an output layer, the output layer comprising models of senones;
- converting the spoken utterance into a plurality of frames;
- providing the plurality of frames to the SI-DNN;
- receiving an output of the SI-DNN based upon the plurality of frames, the output being a probability distribution over the models of the senones;
- and
- adapting the SI-DNN to create an adapted DNN for the particular user based upon the plurality of frames, wherein the adapting of the SI-DNN is undertaken conservatively such that, with respect to the plurality of frames, a deviation between the probability distribution over the models of the senones output by the SI-DNN and a probability distribution output by the adapted DNN is constrained.
Dependencies are inferred from the claim text.