Method of pattern recognition using noise reduction uncertainty

• US 7,460,992 B2
• Filed: 05/16/2006
• Issued: 12/02/2008
• Est. Priority Date: 05/20/2002
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Deng, Li, Droppo, James G., Acero, Alejandro
• Primary Examiner(s): Lerner; Martin
• Application Number:
  US11/435,254
• Publication Number:
  US 20060206325A1
• Time in Patent Office: 931 Days
• Field of Search: 704/226, 704/227, 704/228, 704/233, 704/240, 704/255, 704/256.2, 704/256.3
• US Class Current: 704/226
• 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 acoustic states from a noisy speech signal, the method comprising:

removing noise from a representation of a portion of the noisy speech signal to produce a representation of a portion of a cleaned speech signal, wherein removing noise from a representation of a portion of a noisy speech signal comprises removing noise from a feature vector representing a frame of the noisy speech signal and wherein removing noise from a feature vector comprises;

identifying a mixture component based on the feature vector for the noisy speech signal;

identifying a correction vector and an error value associated with the correction vector based on the identified mixture component; and

using the correction vector, the error value, and the feature vector for the noisy speech signal to identify a feature vector for a frame of the cleaned speech signal;

identifying an uncertainty associated with removing the noise;

using the uncertainty to adjust a probability distribution associated with an acoustic state to form a modified probability distribution; and

applying the representation of a portion of the cleaned speech signal to the modified probability distribution to decode an acoustic state for speech recognition.

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 speech signal comprises a component of a feature vector representing a frame of the noisy speech signal and wherein the representation of a portion of the cleaned speech signal comprises a component of a feature vector representing a frame of the cleaned speech 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 speech 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 identifying an uncertainty comprises utilizing the error value associated with the correction vector to determine the uncertainty.

7. A computer-readable medium having computer-executable instructions for performing steps comprising:

converting a frame of a noisy speech signal into a feature vector comprising at least two components;

removing noise from a component of the feature vector for the noisy speech signal to produce a component of a feature vector for a cleaned speech signal, wherein removing noise comprises;

identifying a correction vector based on the feature vector for the noisy speech signal; and

using the correction vector and the feature vector for the noisy speech signal to form the feature vector for the cleaned speech signal;

identifying an uncertainty associated with removing the noise from the component;

determining a probability component of a probability for a phonetic state by applying the component for the cleaned speech signal to a distribution for the phonetic state defined in part by the uncertainty associated with removing the noise from the component by computing a variance for the distribution using the uncertainty as a term in the computation; and

using the probability component to determine the probability of the phonetic state during speech recognition regardless of the value of the uncertainty.

8. The computer-readable medium of claim 7 wherein defining a probability distribution comprises adding the uncertainty to a variance of a probability distribution.

9. The computer-readable medium of claim 7 wherein identifying a correction vector comprises identifying a mixture component based on the feature vector for the noisy speech signal and selecting a correction vector associated with the mixture component.

10. The computer-readable medium of claim 9 wherein identifying an uncertainty comprises:

identifying an error associated with the mixture component; and

utilizing the error to calculate the uncertainty.