Method of speech recognition using variational inference with switching state space models

• US 6,931,374 B2
• Filed: 04/01/2003
• Issued: 08/16/2005
• Est. Priority Date: 04/01/2003
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Deng, Li, Attias, Hagai, Lee, Leo Jingyu
• Primary Examiner(s): CHAWAN, VIJAY B
• Application Number:
  US10/405,166
• Publication Number:
  US 20040199386A1
• Time in Patent Office: 868 Days
• Field of Search: 700/29, 704/240, 704/232, 704/223, 704/256, 704/252, 704/255
• US Class Current: 704/240
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence

Abstract

A method is developed which includes 1) defining a switching state space model for a continuous valued hidden production-related parameter and the observed speech acoustics, and 2) approximating a posterior probability that provides the likelihood of a sequence of the hidden production-related parameters and a sequence of speech units based on a sequence of observed input values. In approximating the posterior probability, the boundaries of the speech units are not fixed but are optimally determined. Under one embodiment, a mixture of Gaussian approximation is used. In another embodiment, an HMM posterior approximation is used.

Claims

1. A method of estimating a posterior probability for a switching state space model, the posterior probability providing the likelihood of a set of speech units and a set of hidden parameters for a sequence of frames based upon input values associated with the sequence of frames, the method comprising:

approximating the posterior probability in part by multiplying individual hidden parameter probabilities together to form a product of the hidden parameter probabilities, wherein each individual hidden parameter probability provides the probability of a hidden parameter for a frame given a speech unit of the frame and given the input values for the sequence of frames, and by multiplying the product of the hidden parameter probabilities by the product of speech unit probabilities, wherein each speech unit probability provides a probability of a speech unit for a frame; and

adjusting parameters that define the hidden parameter probabilities so that the hidden parameter probabilities provide a better approximation of the posterior probability.

2. The method of claim 1 further comprising adjusting the speech unit probabilities so that the speech unit probabilities provide a better approximation of the posterior probability.

3. The method of claim 1 further comprising using the approximation of the posterior probability to decode input values to identify a sequence of speech units for the sequence of frames.

4. The method of claim 3 wherein using the approximation comprises selecting a speech unit for each frame by selecting the speech unit with the highest speech unit probability for the frame.

5. The method of claim 3 wherein using the approximation comprises applying a duration constraint to the speech units and selecting a best sequence of speech units based on the duration constraint and the speech unit probabilities.

6. The method of claim 1 wherein approximating the posterior probability further comprises multiplying the product of the hidden parameter probabilities by a product of speech unit transition probabilities, wherein each speech unit transition probability provides a probability of a speech unit in a frame given a speech unit in another frame.

7. The method of claim 6 further comprising adjusting the speech unit transition probabilities so that the speech unit transition probabilities provide a better approximation of the posterior probability.

8. The method of claim 6 further comprising using the approximation of the posterior probability to decode input values to identify a sequence of speech units for the sequence of frames.

9. The method of claim 8 wherein using the approximation comprises selecting a sequence of speech units based on the speech unit transition probabilities.

10. The method of claim 9 wherein using the approximation comprises applying a duration constraint to the speech units and selecting a best sequence of speech units based on the duration constraint and the speech unit transition probabilities.

11. The method of claim 1 wherein adjusting the parameters that define the hidden parameter probabilities comprises adjusting the parameters based on parameters of the switching state space model to produce adjusted parameters.

12. The method of claim 11 wherein the input values are generated from a training signal and wherein adjusting the parameters further comprises adjusting the parameters of the switching state space model based on adjusted parameters that define the hidden parameter probabilities to form adjusted switching state space model parameters and adjusting the adjusted parameters that define the hidden parameter probabilities based on the adjusted switching state space model parameters.

13. The method of claim 1 further comprising using the approximation of the posterior probability to identify a sequence of hidden parameters for the sequence of frames.

14. A method of estimating a posterior probability for a switching state space model, the posterior probability providing the likelihood of a set of speech units and a set of hidden parameters for a sequence of frames based upon input values associated with the sequence of frames, the method comprising:

approximating the posterior probability in part by multiplying individual hidden parameter probabilities together to form a product of the hidden parameter probabilities, wherein each individual hidden parameter probability provides the probability of a hidden parameter for a frame given a speech unit of the frame and given the input values for the sequence of frames, and by multiplying the product of the hidden parameter probabilities by a product of speech unit transition probabilities, wherein each speech unit transition probability provides a probability of a speech unit in a frame given a speech unit in another frame; and

adjusting parameters that define the hidden parameter probabilities so that the hidden parameter probabilities provide a better approximation of the posterior probability.

15. The method of claim 14 wherein approximating the posterior probability further comprises multiplying the product of the hidden parameter probabilities by the product of speech unit probabilities, wherein each speech unit probability provides a probability of a speech unit for a frame.

16. The method of claim 15 further comprising adjusting the speech unit probabilities so that the speech unit probabilities provide a better approximation of the posterior probability.

17. The method of claim 15 further comprising using the approximation of the posterior probability to decode input values to identify a sequence of speech units for the sequence of frames.

18. The method of claim 17 wherein using the approximation comprises selecting a speech unit for each frame by selecting the speech unit with the highest speech unit probability for the frame.

19. The method of claim 17 using the approximation comprises applying a duration constraint to the speech units and selecting a best sequence of speech units based on the duration constraint and the speech unit probabilities.

20. The method of claim 14 further comprising adjusting the speech unit transition probabilities so that the speech unit transition probabilities provide a better approximation of the posterior probability.

21. The method of claim 14 further comprising using the approximation of the posterior probability to decode input values to identify a sequence of speech units for the sequence of frames.

22. The method of claim 21 wherein using the approximation comprises selecting a sequence of speech units based on the speech unit transition probabilities.

23. The method of claim 22 wherein using the approximation comprises applying a duration constraint to the speech units and selecting a best sequence of speech units based on the duration constraint and the speech unit transition probabilities.

24. The method of claim 14 wherein adjusting the parameters that define the hidden parameter probabilities comprises adjusting the parameters based on parameters of the switching state space model to produce adjusted parameters.

25. The method of claim 24 wherein the input values are generated from a training signal and wherein adjusting the parameters further comprises adjusting the parameters of the switching state space model based on adjusted parameters that define the hidden parameter probabilities to form adjusted switching state space model parameters and adjusting the adjusted parameters that define the hidden parameter probabilities based on the adjusted switching state space model parameters.

26. The method of claim 14 further comprising using the approximation of the posterior probability to identify a sequence of hidden parameters for the sequence of frames.