Speech models generated using competitive training, asymmetric training, and data boosting

• US 20060287856A1
• Filed: 06/17/2005
• Published: 12/21/2006
• Est. Priority Date: 06/17/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Wu, Jian, He, Xiaodong
• Granted Patent: US 7,693,713 B2
• Time in Patent Office: Days
• US Class Current: 704/256
• CPC Class Codes: G10L 15/063 Training

Abstract

Speech models are trained using one or more of three different training systems. They include competitive training which reduces a distance between a recognized result and a true result, data boosting which divides and weights training data, and asymmetric training which trains different model components differently.

Claims

1. A method of training a speech model, comprising:

obtaining model parameters for the speech model;

processing a known speech input using the speech model with the model parameters to generate a process result;

calculating a distance between a true result and the process result, given the model parameters and the known speech input; and

modifying the model parameters to reduce the distance between the true result and the process result, to obtain a modified model.

2. The method of claim 1 and further comprising:

iterating on the steps of processing, calculating and modifying until the model parameters reach a desired convergence.

3. The method of claim 1 wherein the speech model comprises an acoustic model, and wherein processing a known speech input to generate a process result comprises:

performing speech recognition on acoustic data indicative of a known acoustic input, using the acoustic model with the model parameters, to generate a speech recognition result.

4. The method of claim 3 wherein the true result comprises a true transcription of the acoustic data and wherein calculating a distance comprises:

calculating a true transcription measure indicative of a probability of generating the true transcription, given the acoustic data and the model parameters;

calculating a speech recognition measure indicative of a probability of generating the speech recognition result, given the acoustic data and the model parameters; and

calculating the distance as the difference between the transcription measure and the speech recognition measure.

5. The method of claim 1 wherein the speech model comprises a speech detection model, and wherein processing a known speech input to generate a process result comprises:

performing speech detection on acoustic data indicative of an input signal to generate a detection state output indicative of a decision made by the speech detection model as to whether the input signal represents speech or non-speech.

6. The method of claim 5 wherein the true result comprises a true detection state output indicative of whether the input signal represents speech or non-speech and wherein calculating a distance comprises:

calculating a true detection state measure indicative of a probability that the speech detection model will make a decision that the input signal represents the true detection state output, given the acoustic data and the model parameters;

calculating a speech detection state measure indicative of a probability that the speech detection model will make a decision that the input signal represents the detection state output, given the acoustic data and the model parameters; and

calculating the difference as a difference between the true detection state measure and the speech detection state measure.

7. The method of claim 5 wherein performing speech detection comprises:

generating the detection state output indicative of whether the input signal represents speech, silence, noise or an onset of speech.

8. The method of claim 1 wherein obtaining model parameters for the speech model comprises:

performing maximum likelihood training on training data to obtain an initial model with initial model parameters.

9. A method of training a speech model, comprising:

dividing the speech model into a plurality of sub-model groups based on at least one predetermined criterion;

performing different training on each of the plurality of sub-model groups to obtain a plurality of modified sub-models; and

combining the plurality of modified sub-models to obtain a modified model.

10. The method of claim 9 wherein performing different training comprises:

performing group-specific training on each of the plurality of sub-model groups.

11. The method of claim 9 wherein performing different training comprises:

performing a common training method on the plurality of sub-model groups, using different training settings for each of the plurality of sub-model groups.

12. The method of claim 9 wherein performing different training comprises:

performing different training methods on each of the plurality of sub-model groups.

13. The method of claim 9 wherein performing different training comprises:

performing competitive training, to reduce a distance between a true model result and an actual model result, on at least one of the sub-model groups.

14. The method of claim 9 wherein performing different training comprises:

performing competitive function training on at least one of the sub-model groups to train a competitive function used by the speech model.

15. The method of claim 9 wherein performing different training comprises:

performing data boost training on at least one of the sub-model groups by dividing training data into data groups and weighting each data group, and training at least one sub-model group using the weighted training data.

16. The method of claim 9 wherein dividing the speech model comprises:

dividing the speech model into groups of individual model components so the modified sub-models comprise sets of modified individual model components, and wherein combining comprises forming a superset of the sets of modified individual model components.

17. The method of claim 9 and further comprising:

performing common model training at least either on the speech model before dividing the speech model into the plurality of sub-model groups, or on the modified model after combining the modified sub-models.

18. A method of training a speech model, comprising:

dividing known data into data subgroups based on whether the known data was incorrectly processed by the speech model;

assigning a weight to each of the data subgroups to obtain weighted training data; and

training model parameters in the speech model based on the weighted training data.

19. The method of claim 18 wherein dividing the known data comprises:

dividing the known data based on a type of error made by the speech model in incorrectly processing the known data.

20. The method of claim 18 wherein the speech model comprises one of an acoustic model and a speech detection model.