System and method for training adaptation-specific acoustic models for automatic speech recognition
View Patent ↗Disclosed herein are systems, methods, and computer-readable storage media for training adaptation-specific acoustic models. A system practicing the method receives speech and generates a full size model and a reduced size model, the reduced size model starting with a single distribution for each speech sound in the received speech. The system finds speech segment boundaries in the speech using the full size model and adapts features of the speech data using the reduced size model based on the speech segment boundaries and an overall centroid for each speech sound. The system then recognizes speech using the adapted features of the speech. The model can be a Hidden Markov Model (HMM). The reduced size model can also be of a reduced complexity, such as having fewer mixture components than a model of full complexity. Adapting features of speech can include moving the features closer to an overall feature distribution center.
1. A method comprising:
receiving speech data, to yield received speech data;
generating, via a computing device, a full size adaptation-specific acoustic model and a reduced size adaptation-specific acoustic model, the reduced size adaptation-specific acoustic model starting with a single distribution for each speech sound in the received speech data;
finding speech segment boundaries in the received speech data using the full size adaptation-specific acoustic model;
increasing a size of the reduced size adaptation-specific acoustic model using the full size adaptation-specific acoustic model to meet a desired performance level, to yield a modified reduced size adaptation-specific acoustic model;
after the finding of the speech segment boundaries, adapting features of the received speech data using the modified reduced size adaptation-specific acoustic model based on the speech segment boundaries and an overall centroid for each speech sound, to yield adapted features; and
recognizing the received speech data based on the adapted features.
2. The method of claim 1 , wherein the full size adaptation-specific acoustic model and the reduced size adaptation-specific model are hidden markov models.
3. The method of claim 1 , wherein the reduced size adaptation-specific acoustic model is a model of reduced complexity.
4. The method of claim 3 , wherein the reduced size adaptation-specific acoustic model has less mixture components than a model of full complexity.
5. The method of claim 3 , wherein the reduced size adaptation-specific acoustic model has less mixture components per state than a model of full complexity.
6. The method of claim 1 , wherein adapting features of the received speech data further comprises moving the features closer to a center of an overall feature distribution.
7. The method of claim 1 , further comprising recursively training a recognition model based on the adapted features, to yield a recursively trained model.
8. The method of claim 7 , further comprising adapting speech of a new speaker at recognition time using the recursively trained model.
9. A system comprising:
a processor; and
a computer-readable storage device having instructions stored which, when executed by the processor, result in the processor performing operations comprising:
receiving speech data, to yield received speech data;
generating a full size adaptation-specific acoustic model and a reduced size adaptation-specific acoustic model, the reduced size adaptation-specific acoustic model starting with a single distribution for each speech sound in the received speech data;
finding speech segment boundaries in the received speech data using the full size adaptation-specific acoustic model;
increasing a size of the reduced size adaptation-specific acoustic model using the full size adaptation-specific acoustic model to meet a desired performance level, to yield a modified reduced size adaptation-specific acoustic model;
after the finding of the speech segment boundaries, adapting features of the received speech data using the modified reduced size adaptation-specific acoustic model based on the speech segment boundaries and an overall centroid for each speech sound, to yield adapted features; and
recognizing the received speech data based on the adapted features.
10. The system of claim 9 , wherein the full size adaptation-specific acoustic model and the reduced size adaptation-specific model are hidden markov models.
11. The system of claim 9 , wherein the reduced size adaptation-specific acoustic model is a model of reduced complexity.
12. The system of claim 11 , wherein the reduced size adaptation-specific acoustic model has less mixture components than a model of full complexity.
13. The system of claim 11 , wherein the reduced size adaptation-specific acoustic model has less mixture components per state than a model of full complexity.
14. The system of claim 9 , wherein the adapting of the features of the received speech data further comprises moving the features closer to a center of an overall feature distribution.
15. The system of claim 9 , the computer-readable storage device having additional instructions stored which result in the operations further comprising recursively training a recognition model based on the adapted features, to yield a recursively trained model.
16. The system of claim 15 , the computer-readable storage device having additional instructions stored which result in the operations further comprising adapting speech of a new speaker at recognition time using the recursively trained model.
17. A computer-readable storage device having instructions stored which, when executed by a computing device, cause the computing device to perform operations comprising:
receiving speech data, to yield received speech data;
generating a full size adaptation-specific acoustic model and a reduced size adaptation-specific acoustic model, the reduced size adaptation-specific acoustic model starting with a single distribution for each speech sound in the received speech data;
finding speech segment boundaries in the received speech data using the full size adaptation-specific acoustic model;
increasing a size of the reduced size adaptation-specific acoustic model using the full size adaptation-specific acoustic model to meet a desired performance level, to yield a modified reduced size adaptation-specific acoustic model;
after the finding of the speech segment boundaries, adapting features of the received speech data using the modified reduced size adaptation-specific acoustic model based on the speech segment boundaries and an overall centroid for each speech sound, to yield adapted features; and
recognizing the received speech data based on the adapted features.
18. The computer-readable storage device of claim 17 , wherein the full size adaptation-specific acoustic model and the reduced size adaptation-specific model are hidden markov models.
19. The computer-readable storage device of claim 18 , wherein the reduced size adaptation-specific acoustic model is a model of reduced complexity.
20. The computer-readable storage device of claim 19 , wherein the reduced size adaptation-specific acoustic model has less mixture components than a model of full complexity.