Training system for a speech recognition application

• US 20070219798A1
• Filed: 03/16/2006
• Published: 09/20/2007
• Est. Priority Date: 03/16/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Wang, Ye-Yi, Acero, Alex, Lee, John
• Granted Patent: US 7,813,926 B2
• Time in Patent Office: Days
• US Class Current: 704/257
• CPC Class Codes: G10L 15/063 Training

Abstract

A training system for a speech recognition application is disclosed. In embodiments described, the training system is used to train a classification model or language model. The classification model is trained using an adaptive language model generated by an iterative training process. In embodiments described, the training data is recognized by the speech recognition component and the recognized text is used to create the adaptive language model which is used for speech recognition in a following training iteration.

Claims

1. A training system implementable on a computer readable medium comprising:

a training component configured to invoke a speech recognition component to recognize training utterances in each of a plurality of training iterations using an adaptive language model for one or more of the plurality of training iterations and output recognized text; and

a classification model training tool configured to generate a classification model based upon the recognized text from the training component.

2. The training system of claim 1 wherein the training component includes computer implementable instructions to recognize a first set of training utterances based upon a general language model and recognize a second set of training utterances using the adaptive language model generated using the first set of training utterances.

3. The training system of claim 1 wherein the training component includes instructions to implement the plurality of training iterations for a plurality of training data sets using an unsupervised language model adaptation.

4. The training system of claim 1 wherein the training component includes instructions to perform repeat training iterations for one or more of a plurality of training data sets using the adaptive language model from a previous training iteration.

5. The training system of claim 1 wherein the training component is configured to designate an iterative language model for use by the speech recognition component and in a first training iteration a general language model is designated as the iterative language model and in a subsequent training iteration the adaptive language model is designated as the iterative language model for speech recognition.

6. The training system of claim 5 wherein the adaptive language model is derived based upon an interpolation of the iterative language model and a domain-specific language model generated by a language model training tool in the plurality of training iterations.

7. The training system of claim 1 and further comprising a testing component that outputs one or more of a classification error rate, word error rate or conditional entropy and the training component being configured to evaluate whether to perform additional training iterations based upon the classification error rate, the word error rate or the conditional entropy.

8. The training system of claim 1 wherein the training component includes computer implementable instructions to recognize utterances in n-training data sets in one or more of the plurality of training iterations and uses n adaptive language models to recognize utterances in the n-training data sets in subsequent iterations.

9. A method implementable on a computer readable medium comprising the steps of:

recognizing speech utterances in one or more sets of training data and outputting recognized text using an adaptive language model; and

generating a classification model using the recognized text and a class annotation from the one or more sets of training data.

10. The method of claim 9 and further comprising the step of:

generating the adaptive language model based upon the recognized text in one iteration and using the adaptive language model for recognition of the speech utterances in the one or more sets of training data in another training iteration.

11. The method of claim 9 and further comprising the step of:

recognizing the speech utterances in a first set of training data in a first training iteration and generating the adaptive language model based upon the recognized text of the first training iteration; and

recognizing the speech utterances in a second set of training data in a second training iteration using the adaptive language model generated in the first training iteration.

12. The method of claim 9 and further comprising the steps of:

recognizing the speech utterances for n-training data sets using a general language model and generating n-adaptive language models for the n-training data sets; and

repeating recognition of each n training data set of training data in subsequent training iterations using the adaptive language model obtained with n−

1 training data sets from previous training iterations.

13. The method of claim 9 and further comprising the steps of:

recognizing the speech utterances for a first set of training data in a first training iteration using a general language model; and

interpolating a domain-specific language model generated from the recognized text of the first set of training data and the general language model to create the adaptive language model.

14. The method of claim 13 and further comprising the steps of:

designating the adaptive language model as an iterative language model;

recognizing a second set of training data in a second training iteration using the iterative language model; and

interpolating between the iterative language model and the domain-specific language model generated from recognized text for the second set of training data to generate the adaptive language model for a subsequent training iteration.

15. The method of claim 9 and further comprising the steps of:

testing classification or recognition of a development data set using the adaptive language model and the classification model; and

repeating the step of recognizing the speech utterances for the set of training data based upon classification or recognition improvement.

16. The method of claim 15 wherein the step of testing classification or recognition of the development data set includes the step of:

calculating a word error rate or classification error rate for recognition or classification of the development data set.

17. The method of claim 15 wherein the step of testing classification or recognition comprises:

calculating a conditional entropy for recognition of the development data set.

18. A computer readable medium storing computer readable instructions which, when executed by a computer perform the steps of:

recognizing training utterances in each of a plurality of training data sets and generating an adaptive language model based upon text recognition of the training utterances in the plurality of training data sets; and

repeating recognition of the training utterances in one or more of the plurality of training data sets using the adaptive language model of a previous training iteration.

19. The computer readable medium of claim 18 and further comprising instructions which when executed generate a classification model based upon recognition of the training utterances from the plurality of training data sets and class annotations for the training utterances.

20. The computer readable medium of claim 18 and further comprising instructions which when executed perform the steps of:

recognizing the training utterances in a first set of training data and generating the adaptive language model;

designating the adaptive language model as the iterative language model;

recognizing the training utterances in a second set of training data using the iterative language model; and

repeating recognition of each set of training data using the adaptive language model generated by a previous training iteration.