METHOD OF PATTERN RECOGNITION USING NOISE REDUCTION UNCERTAINTY

• US 20080281591A1
• Filed: 07/25/2008
• Published: 11/13/2008
• Est. Priority Date: 05/20/2002
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Deng, Li, Droppo, James G., Acero, Alejandro
• Granted Patent: US 7,769,582 B2
• Time in Patent Office: Days
• US Class Current: 704/233
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 21/0208 Noise filtering

Abstract

A method and apparatus are provided for using the uncertainty of a noise-removal process during pattern recognition. In particular, noise is removed from a representation of a portion of a noisy signal to produce a representation of a cleaned signal. In the meantime, an uncertainty associated with the noise removal is computed and is used with the representation of the cleaned signal to modify a probability for a phonetic state in the recognition system. In particular embodiments, the uncertainty is used to modify a probability distribution, by increasing the variance in each Gaussian distribution by the amount equal to the estimated variance of the cleaned signal, which is used in decoding the phonetic state sequence in a pattern recognition task.

Claims

1. A method of recognizing pattern states from a noisy signal, the method comprising:

removing noise from a representation of a portion of the noisy signal to produce a representation of a portion of a cleaned signal by applying a correction vector for a mixture component;

identifying an uncertainty associated with removing the noise by applying a variance for a prior probability distribution of a clean signal and a variance associated with selecting the correction vector for the mixture component to a function;

using the uncertainty to adjust a probability distribution to form a modified probability distribution for the mixture component; and

applying the representation of a portion of the cleaned signal to the modified probability distribution to decode a pattern state.

2. The method of claim 1 wherein adjusting a probability distribution comprises adding the uncertainty to a variance of the probability distribution to form the modified probability distribution.

3. The method of claim 1 wherein the representation of a portion of the noisy signal comprises a component of a feature vector representing a frame of the noisy signal and wherein the representation of a portion of the cleaned signal comprises a component of a feature vector representing a frame of the cleaned signal.

4. The method of claim 3 wherein identifying an uncertainty comprises identifying an uncertainty associated with removing noise from the component of the feature vector for the noisy signal to form the component of the feature vector for the cleaned signal.

5. The method of claim 4 wherein using the uncertainty to adjust a probability distribution comprises using the uncertainty to adjust a probability distribution associated with the component of the feature vector.

6. The method of claim 1 wherein applying a variance for a prior probability distribution of a clean signal and a variance associated with selecting the correction vector for the mixture component to a function comprises multiplying the variances together and dividing the product by a difference between the variances.

7. The method of claim 1 wherein the variance for the prior probability distribution is determined by taking a sum over all possible mixture components.

8. The method of claim 1 wherein the variance associated with selecting the correction vector for the mixture component is calculated before receiving the representation of the portion of the noisy signal.

9. A computer-readable storage medium having stored thereon computer-executable instructions that when executed by a process cause the process to perform steps comprising:

converting a frame of a noisy signal into a feature vector;

removing noise from the feature vector for the noisy signal to produce a feature vector for a cleaned signal by applying a correction vector to the feature vector for the noisy signal;

identifying an uncertainty associated with removing the noise from the feature vector for the noisy signal by calculating a value for the uncertainty using an error associated with the correction vector and a variance for a prior probability distribution for clean signal feature vectors; and

determining a probability for a phonetic state by applying the feature vector for the cleaned signal to a probability distribution defined in part by the uncertainty associated with removing the noise from the feature vector for the noisy signal.

10. The computer-readable storage medium of claim 9 wherein defining a probability distribution comprises adding the uncertainty to a variance of a probability distribution to form a variance.

11. The computer-readable storage medium of claim 9 wherein applying a correction vector comprises identifying a mixture component based on the feature vector for the noisy signal and selecting a correction vector associated with the mixture component.

12. The computer-readable storage medium of claim 9 wherein the variance for a prior probability distribution for clean signal feature vectors is determined by taking a summation across a plurality of mixture components.

13. The computer-readable storage medium of claim 9 wherein the error associated with the correction vector is determined before converting a frame of a noisy signal into a feature vector.

14. A computer-readable storage medium having stored thereon computer-executable instructions that when executed by a processor cause the processor to perform steps comprising:

identifying an uncertainty value for a noise removal process used to form a noise-reduced value from a noisy value by applying to a function an error associated with a correction vector used in the noise removal process and a variance for clean values; and

utilizing the uncertainty value to define a variance for a probability distribution that provides a probability that is proportional to a probability of the noisy value given a mixture component associated with a phonetic state.

15. The computer-readable storage medium of claim 14 wherein identifying an uncertainty value comprises identifying one of a plurality of mixture components and selecting an uncertainty value associated with the identified mixture component.

16. The computer-readable storage medium of claim 15 wherein the variance for clean values is computed based on a sum over the plurality of mixture components.

17. The computer-readable storage medium of claim 16 wherein probability provided by the probability distribution is determined by applying the noise-reduced value to the probability distribution.

18. The computer-readable storage medium of claim 14 wherein the step of identifying an uncertainty value is performed for each frame of a noisy signal.

19. The computer-readable storage medium of claim 14 wherein applying to a function an error associated with a correction vector used in the noise removal process and a variance for clean values comprises forming a product of the error and the variance and dividing the product by a difference between the variance and the error.

20. The computer-readable storage medium of claim 14 wherein utilizing the uncertainty value to define a variance comprises summing the uncertainty value with a variance for a probability distribution for clean values for a mixture component.