Removing noise from feature vectors

• US 7,451,083 B2
• Filed: 07/20/2005
• Issued: 11/11/2008
• Est. Priority Date: 03/20/2001
• Status: Expired due to Term

• Current Assignee: SZ DJI Technology Co., Ltd. (d/b/a DJI) (iFlight Technology Company Limited)
• Original Assignee: Microsoft Corporation
• Inventors: Deng, Li, Acero, Alejandro, Frey, Brendan J.
• Primary Examiner(s): Wozniak; James S
• Application Number:
  US11/185,522
• Publication Number:
  US 20050256706A1
• Time in Patent Office: 1,210 Days
• Field of Search: 704/233, 704239-240, 704/236, 704/231, 704243-244, 704/251
• US Class Current: 704/233
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• 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

Abstract

A method and computer-readable medium are provided for identifying clean signal feature vectors from noisy signal feature vectors. One aspect of the invention includes using an iterative approach to identify the clean signal feature vector. Another aspect of the invention includes using the variance of a set of noise feature vectors and/or channel distortion feature vectors when identifying the clean signal feature vectors.

Claims

1. A method comprising:

identifying a mixture of distributions that provide prior probabilities for combinations of clean signal feature vectors and obscuring feature vectors, the mixture of distributions comprising mixture components each comprising a mean and variance;

determining an observation variance for an observation probability that provides the probability of a noisy signal feature vector given a clean signal feature vector and at least one obscuring feature vector;

calculating a mean for a posterior probability distribution by applying a mean and a variance of a mixture component of the mixture of distributions that provide prior probabilities and the observation variance to a function;

using the mean of the posterior probability distribution to identify the clean signal feature vector; and

using the clean signal feature to identify a word during speech recognition.

2. The method of claim 1 wherein determining an observation variance comprises using a closed-form expression to determine the variance.

3. The method of claim 1 wherein determining an observation variance comprises:

generating a feature vector from a clean signal;

generating a feature vector from a noise signal;

generating an observed noisy signal feature vector from a noisy signal that is the combination of at least the clean signal and the noise signal;

calculating a calculated noisy signal feature vector from the feature vector from the clean signal and the feature vector from the noise signal;

comparing the observed noisy signal feature vector to the calculated noisy signal feature vector to determine the variance.

4. The method of claim 3 wherein determining an observation variance further comprises determining a feature vector for channel distortion and wherein calculating a calculated noisy signal feature vector further comprises calculating the calculated noisy signal feature vector in part from the feature vector for channel distortion.

5. The method of claim 1 wherein identifying a mixture of distributions that provide prior probabilities comprises identifying a mixture of distributions for the obscuring feature vectors and a mixture of distributions for the clean signal feature vectors.

6. A method comprising:

accessing a prior mean for a distribution of training feature vectors that represents a prior probability of combinations of training feature vectors wherein the prior mean of the distribution of training feature vectors comprises a prior mean vector comprising elements representing a mean of a mixture component of a mixture of distributions for clean signal feature vectors and elements representing a mean of a mixture component of a mixture of distributions for noise training feature vectors;

setting an initial value for a posterior mean vector comprising elements for a component of a clean signal feature vector and elements for a component of a noise feature vector;

determining a revised value for the posterior mean vector based in part on the initial value for the posterior mean vector, a difference between the prior mean vector and the initial value for the posterior mean vector, and a noisy signal feature vector, wherein the noisy signal feature vector represents a noisy signal that is a combination of a clean signal with at least one of a group comprising a noise signal and channel distortion;

determining whether to accept the revised value as a final value for the posterior mean vector;

using the final value for the posterior mean vector to identify a clean signal feature vector; and

using the clean signal feature vector to recognize a word during speech recognition.

7. The method of claim 6 wherein determining whether to accept the revised value as a final value comprises:

deciding not to accept the revised value as the final value;

determining a further revised value for the posterior mean vector based in part on the revised value for the posterior mean vector, the difference between the prior mean vector and the revised value for the posterior mean vector, and the noisy signal feature vector; and

determining whether to accept the further revised value as a final value for the posterior mean vector.

8. The method of claim 6 wherein the steps of the method are repeated for a plurality of distributions of training feature vectors to generate a final value for a posterior mean vector for a plurality of components of the clean signal feature vector.

9. The method of claim 6 wherein determining a revised value comprises determining a revised value based in part on a variance associated with the noisy signal feature vector.

10. The method of claim 6 wherein the distribution of training feature vectors is described in part by a variance.

11. The method of claim 6 wherein accessing a prior mean for a distribution further comprises accessing a prior mean vector comprising elements representing a mean of a mixture component of a mixture of distributions for channel distortion training feature vectors.

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

accessing a noisy signal feature vector that represents a noisy signal which is a combination of a clean signal with at least one of a group comprising noise and channel distortion;

accessing prior mean vector for at least one distribution of training feature vectors wherein accessing a prior mean vector comprises accessing a prior mean vector comprising elements representing a mean vector for a distribution of noise training feature vectors described in part by a variance, and elements representing a mean vector for a distribution of clean signal training feature vectors;

identifying an initial value for a posterior mean vector comprising elements for a clean signal feature vector and elements for a noise feature vector; and

performing iterations to identify a final value for the posterior mean vector, each iteration performing a calculation based on the noisy signal feature vector, the difference between the prior mean vector and a current value for the posterior mean vector, and the current value for the posterior mean vector, the current value for the posterior mean vector being updated with each iteration;

using the final value for the posterior mean vector to identify a clean signal feature vector; and

using the clean signal feature vector in speech recognition to recognize a word.

13. The computer-readable storage medium of claim 12 wherein performing a calculation at each iteration further comprises performing the calculation based on a variance for the noisy signal feature vector.

14. The computer-readable storage medium of claim 12 wherein the distribution of noise training feature vectors is a mixture component of a mixture of distributions for noise training feature vectors.

15. The computer-readable storage medium of claim 12 wherein the prior mean vector further comprises elements representing a mean vector for a distribution of channel distortion training feature vectors described in part by a variance.

16. The computer-readable storage medium of claim 15 wherein the distribution of channel distortion training feature vectors is a mixture component of a mixture of distributions for channel distortion training feature vectors.

17. The computer-readable storage medium of claim 16 wherein a first prior mean vector comprises elements representing a mean vector for a distribution of channel distortion training feature vectors associated with a first channel and a second prior mean vector comprises elements representing a mean vector for a distribution of channel distortion training feature vectors associated with a second channel.