Multi-Class Constrained Maximum Likelihood Linear Regression

• US 20090024390A1
• Filed: 05/02/2008
• Published: 01/22/2009
• Est. Priority Date: 05/04/2007
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Deshmukh, Neeraj, Zhan, Puming
• Granted Patent: US 8,386,254 B2
• Time in Patent Office: Days
• US Class Current: 704/236
• CPC Class Codes: G10L 15/07 to the speaker

Abstract

A method of speech recognition converts an unknown speech input into a stream of representative features. The feature stream is transformed based on speaker dependent adaptation of multi-class feature models. Then automatic speech recognition is used to compare the transformed feature stream to multi-class speaker independent acoustic models to generate an output representative of the unknown speech input.

Claims

1. A method of speech recognition comprising:

converting an unknown speech input into a stream of representative features;

transforming the feature stream based on speaker dependent adaptation of multi-class feature models; and

using automatic speech recognition to compare the transformed feature stream to multi-class speaker independent acoustic models to generate an output representative of the unknown speech input.

2. A method according to claim 1, wherein the adaptation is based on constrained maximum likelihood linear regression (cMLLR).

3. A method according to claim 2, wherein the adaptation based on cMLLR includes use of relevance smoothing.

4. A method according to claim 1, wherein the converting an unknown speech input is based on a speaker dependent front end.

5. A method according to claim 1, wherein the transforming the feature stream includes using speaker adaptive training (SAT).

6. A method according to claim 1, wherein the transforming the feature stream includes using online unsupervised feature space adaptation (OUFA).

7. A method according to claim 1, wherein the converting an unknown speech input and the transforming the feature stream are performed by a local client, and the using automatic speech recognition is performed by a remote server.

8. A method according to claim 1, wherein the multi-class independent acoustic models include hidden Markov models.

9. A method according to claim 8, wherein the hidden Markov models use Gaussian mixture components.

10. A method according to claim 1, wherein the speech recognition uses scaled integer arithmetic.

11. A system for speech recognition comprising:

means for converting an unknown speech input into a stream of representative features;

means for transforming the feature stream based on speaker dependent adaptation of multi-class feature models; and

means for using automatic speech recognition to compare the transformed feature stream to multi-class speaker independent acoustic models to generate an output representative of the unknown speech input.

12. A system according to claim 11, wherein the adaptation in the means for transforming is based on constrained maximum likelihood linear regression (cMLLR).

13. A system according to claim 12, wherein the adaptation based on cMLLR includes use of relevance smoothing.

14. A system according to claim 11, wherein the means for converting an unknown speech input is based on a speaker dependent front end.

15. A system according to claim 11, wherein the means for transforming the feature stream includes means for using speaker adaptive training (SAT).

16. A system according to claim 11, wherein the means for transforming the feature stream includes means for using online unsupervised feature space adaptation (OUFA).

17. A system according to claim 11, wherein the means for converting an unknown speech input and the means for transforming the feature stream are at a local client, and the means for using automatic speech recognition is at a remote server.

18. A system according to claim 11, wherein the multi-class independent acoustic models include hidden Markov models.

19. A system according to claim 18, wherein the hidden Markov models use Gaussian mixture components.

20. A system according to claim 11, wherein the means for using speech recognition includes means for using scaled integer arithmetic.