On-line parametric histogram normalization for noise robust speech recognition

• US 20030204398A1
• Filed: 04/30/2002
• Published: 10/30/2003
• Est. Priority Date: 04/30/2002
• Status: Active Grant

• Current Assignee: RPX Corporation
• Original Assignee: Nokia Corporation
• Inventors: Kiss, Imre, Haverinen, Hemmo
• Granted Patent: US 7,197,456 B2
• Time in Patent Office: Days
• 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
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band

Abstract

A method for improving noise robustness in speech recognition, wherein a front-end is used for extracting speech feature from an input speech and for providing a plurality of scaled spectral coefficients. The histogram of the scaled spectral coefficients is normalized to the histogram of a training set using Gaussian approximations. The normalized spectral coefficients are then converted into a set of cepstrum coefficients by a decorrelation module and further subjected to ceptral domain feature-vector normalization.

Claims

1. A method of improving noise robustness in a speech recognition system, the system including a front-end for extracting speech features from an input speech and a back-end for speech recognition based on the extracted features, wherein the front-end comprises:

means, responsive to the input speech, for providing data indicative of the input speech at a plurality of time instants;

means, responsive to the data segments, for spectrally converting the data segments into a plurality of spectral coefficients having a related probability distribution of values for providing spectral data indicative of the spectral coefficients; and

means, responsive to the spectral data, for performing decorrelation conversion on the spectral coefficients for providing the extracted features, characterized by

obtaining a parametric representation of the probability distribution of values of the spectral coefficients;

modifying the parametric representation based on one or more reference values; and

adjusting at least one of the spectral coefficients based on the modified parametric representation for changing the spectral data prior to the decorrelation conversion.

2. The method of claim 1, wherein a plurality of spectral coefficients of a training speech are used for matching, said method further characterized in that

said one or more reference values include a mean value of the spectral coefficients of the training speech.

3. The method of claim 2, further characterized in that

said one or more reference values further include a standard deviation of the spectral coefficients of the training speech.

4. The method of claim 1, further characterized in that

the parametric representation comprises a mean value of the probability distribution of values of the spectral coefficients.

5. The method of claim 1, further characterized in that

the parametric representation comprises a standard deviation of the probability distribution of values of the spectral coefficients.

6. The method of claim 1, further characterized in that

the parametric representation is obtained based on a Gaussian approximation.

7. The method of claim 3, wherein the spectral coefficients of the training speech have a further probability distribution of values, said method further characterized in that

the mean value and the standard deviation are obtained based on a Gaussian approximation of the further probability distribution.

8. A speech recognition front-end for use in a speech recognition system having a back-end, the front end extracting speech features from an input speech so as to allow the back-end to recognize the input speech based on the extracted features, the front-end comprising:

means, responsive to the input speech, for providing data indicative of the input speech at a plurality of time instants;

means for spectrally converting the data into a plurality of spectral coefficients having a related probability distribution of values for providing spectral data indicative of the spectral coefficients; and

means for performing decorrelation conversion on the spectral coefficients for providing the extracted features to the back-end, characterized by

means, responsive to the spectral coefficients, for obtaining a parametric representation of the probability distribution of values of the spectral, for modifying the parametric representation based on one or more reference values, and for adjusting at least one of the spectral coefficients based on the modified parametric representation for changing the spectral data prior to the performing of the decorrelation conversion.

9. The front-end of claim 8, wherein a plurality of spectral coefficients of a training speech are used for matching, said system further characterized in that

said one or more reference values include a mean value of the spectral coefficients of the training speech.

10. The front-end of claim 9, further characterized in that

said one or more reference values further include a standard deviation of the spectral coefficients of the training speech.

11. The front-end of claim 8, further characterized in that

the parametric representation comprises a mean value of the probability distribution of values of the spectral coefficients.

12. The front-end of claim 8, further characterized in that

the parametric representation comprises a standard deviation of the probability distribution of values of the spectral coefficients.

13. The front-end of claim 8, further characterized in that

the parametric representation is obtained based on a Gaussian approximation.

14. The front-end of claim 10, the spectral coefficients of the training speech have a further probability distribution of values, said front-end further characterized in that

the mean value and the standard deviation are obtained based on a Gaussian approximation of the further probability distribution.

15. A network element in a communication system including a back-end for receiving speech data from the network element, the network element comprising:

a voice input device to receive input speech; and

a front-end, responsive to the input speech, for extracting speech features from the input speech for providing speech data indicative of the speech features so as to allow the back-end to recognize the input speech based on the speech features, wherein the front-end comprises;

means, responsive to the input speech, for providing data indicative of the input speech at a plurality of time instants;

means for spectrally converting the data into a plurality of spectral coefficients for providing spectral data indicative of the spectral coefficients having a related probability distribution of values; and

means for performing decorrelation conversion on the spectral coefficients for providing the extracted features, said network element characterized in that

the front-end further comprises means, responsive to the spectral coefficients, for obtaining a parametric representation of the probability distribution of values of the spectral coefficients, for modifying the parametric representation based on one or more reference values, and for adjusting at least one of the spectral coefficients based on the modified parametric representation for changing the spectral data prior to the performing of the decorrelation conversion.

16. The network element of claim 15, wherein a plurality of spectral coefficients of a training speech are used for matching, said network element further characterized in that

said one or more reference values include a mean value of the spectral coefficients of the training speech.

17. The network element of claim 16, further characterized in that

said one or more reference values further include a standard deviation of the spectral coefficients of the training speech.

18. The network element of claim 15, further characterized in that

the parametric representation comprises a mean value of the probability distribution of values of the spectral coefficients.

19. The network element of claim 15, further characterized in that

the parametric representation comprises a standard deviation of the probability distribution of values of the spectral coefficients.

20. The network element of claim 15, further characterized in that

the parametric representation is obtained based on a Gaussian approximation.

21. The network element of claim 16, wherein the spectral coefficients of the training speech have a further probability distribution of value, said method further characterized in that

the mean value and the standard deviation are obtained based on a Gaussian approximation of the further probability distribution.

22. A computer program for use in a speech recognition front-end for extracting speech features from an input speech so as to allow a speech recognition back-end to recognize the input speech based on the extracted features, wherein the front-end comprises:

means, responsive to the input speech, for providing data indicative of the input speech at a plurality of time instants;

means for spectrally converting the data into a plurality of spectral coefficients having a related probability distribution of values for providing spectral data indicative of the spectral coefficients; and

means for performing decorrelation conversion on the spectral coefficients for providing the extracted feature, said computer program characterized by

an algorithm for generating a parametric representation of the probability distribution of values of the spectral coefficients, for modifying the parametric representation based on one or more reference values, and for adjusting at least one of the spectral coefficients based on the modified parametric representation for changing the spectral data prior to the performing of the decorrelation conversion.

23. The computer program of claim 22, wherein a plurality of spectral coefficients of a training speech are used for matching, said computer program further characterized in that

said one or more reference values include a mean value of the spectral coefficients of the training speech.

24. The computer of program of claim 23, further characterized in that

said one or more reference values include a standard deviation of the spectral coefficients of the training speech.

25. The computer program of claim 22, further characterized in that

the parametric representation comprises a mean value of the probability distribution of values of the spectral coefficients.

26. The computer program of claim 22, further characterized in that

the parametric representation comprises a standard deviation of the probability distribution of values of the spectral coefficients.

27. The computer program of claim 22, further characterized in that

the parametric representation is obtained based on a Gaussian approximation.

28. The computer program of claim 24, wherein the coefficients of the training speech has a further probability distribution of values, said computer program further characterized in that

the mean value and the standard deviation are obtained from a Gaussian approximation of the further probability distribution.