COMPUTATIONALLY EFFICIENT METHOD AND APPARATUS FOR SPEAKER RECOGNITION

• US 20040111261A1
• Filed: 12/10/2002
• Published: 06/10/2004
• Est. Priority Date: 12/10/2002
• Status: Active Grant

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Ramaswamy, Ganesh N., Chaudhari, Upendra V., Zilca, Ran
• Granted Patent: US 6,772,119 B2
• Time in Patent Office: Days
• US Class Current: 704/236
• CPC Class Codes: G10L 17/08 Use of distortion metrics or a particular distance between probe pattern and reference templates

Abstract

A speaker recognition technique is provided that can operate within the memory and processing constraints of existing portable computing devices. A smaller memory footprint and computational efficiency are achieved using single Gaussian models for each enrolled speaker. During enrollment, features are extracted from one or more enrollment utterances from each enrolled speaker, to generate a target speaker model based on a sample covariance matrix. During a recognition phase, features are extracted from one or more test utterances to generate a test utterance model that is also based on the sample covariance matrix. A sphericity ratio is computed that compares the test utterance model to the target speaker model, as well as a background model. The sphericity ratio indicates how similar test utterance speech is to the speech used when the user was enrolled, as represented by the target speaker model, and how dissimilar the test utterance speech is from the background model.

Claims

1. A method for recognizing a speaker, comprising the steps of:

extracting features from one or more enrollment utterances to generate a target speaker model based on a sample covariance matrix generated from said extracted enrollment features;

extracting features from one or more test utterances to generate a test utterance model based on a sample covariance matrix generated from said extracted test features; and

computing a sphericity ratio to recognize said speaker, said sphericity ratio comparing said test utterance model to said target speaker model and to a background model.

2. The method of claim 1, wherein said one or more enrollment utterances are obtained in an acoustic environment that is substantially matched to an acoustic environment of said one or more test utterances.

3. The method of claim 2, wherein said acoustic environment is associated with a portable computing device.

4. The method of claim 1, wherein said sphericity ratio is used to recognize the identity of said speaker by comparing said sphericity ratio of a given speaker model to said sphericity ratios obtained against other speaker models.

5. The method of claim 1, further comprising the step of receiving an asserted identity from said speaker, and wherein said sphericity ratio is used to recognize the identity of said speaker by applying a threshold to said sphericity ratio to determine whether said speaker is a person associated with said asserted identity.

6. The method of claim 1, wherein said sphericity ratio may be expressed as follows:

Score=-trace(CU inv(CS))/trace(CU inv(CB));

where CS is a sample covariance matrix of the features extracted from the one or more enrollment utterances;

CB is the sample covariance matrix of the features extracted from background data, and CU is the sample covariance matrix of the features extracted from the one or more test utterances.

7. The method of claim 6, wherein said expression trace(CU inv(CS)) evaluates how dissimilar said one or more test utterances are from said one or more enrollment utterances.

8. The method of claim 6, wherein said expression trace(CU inv(CB)) evaluates how dissimilar said one or more test utterances are from said background model.

9. The method of claim 1, wherein said background model is obtained from utterances of a plurality of speakers.

10. A method for recognizing a speaker, comprising the steps of:

extracting features from one or more enrollment utterances to generate a target speaker model based on a sample covariance matrix generated from said extracted enrollment features;

extracting features from one or more test utterances to generate a test utterance model based on a sample covariance matrix generated from said extracted test features; and

computing a normalized similarity score that compares said test utterance model to said target speaker model and to a background model.

11. The method of claim 10, wherein said one or more enrollment utterances are obtained in an acoustic environment that is substantially matched to an acoustic environment of said one or more test utterances.

12. The method of claim 11, wherein said acoustic environment is associated with a portable computing device.

13. The method of claim 10, wherein said normalized similarity score is used to recognize the identity of said speaker by comparing said normalized similarity score of a given speaker model to said normalized similarity scores obtained against other speaker models.

14. The method of claim 10, further comprising the step of receiving an asserted identity from said speaker, and wherein said normalized similarity score is used to recognize the identity of said speaker by applying a threshold to said normalized similarity score to determine whether said speaker is a person associated with said asserted identity.

15. The method of claim 10, wherein said normalized similarity score may be expressed as follows:

Score=-trace(CU inv(CS))/trace(CU inv(CB));

where CS is a sample covariance matrix of the features extracted from the one or more enrollment utterances;

CB is the sample covariance matrix of the features extracted from background data, and CU is the sample covariance matrix of the features extracted from the one or more test utterances.

16. An apparatus for recognizing a speaker, comprising:

a memory; and

at least one processor, coupled to the memory, operative to;

extract features from one or more enrollment utterances to generate a target speaker model based on a sample covariance matrix generated from said extracted enrollment features;

extract features from one or more test utterances to generate a test utterance model based on a sample covariance matrix generated from said extracted test features; and

compute a sphericity ratio to recognize said speaker, said sphericity ratio comparing said test utterance model to said target speaker model and to a background model.

17. The apparatus of claim 16, wherein said one or more enrollment utterances are obtained in an acoustic environment that is substantially matched to an acoustic environment of said one or more test utterances.

18. The apparatus of claim 17, wherein said acoustic environment is associated with a portable computing device.

19. The apparatus of claim 16, wherein said sphericity ratio is used to recognize the identity of said speaker by comparing said sphericity ratio of a given speaker model to said sphericity ratios obtained against other speaker models.

20. The apparatus of claim 16, further comprising the step of receiving an asserted identity from said speaker, and wherein said sphericity ratio is used to recognize the identity of said speaker by applying a threshold to said sphericity ratio to determine whether said speaker is a person associated with said asserted identity.

21. The apparatus of claim 16, wherein said sphericity ratio may be expressed as follows:

Score=-trace(CU inv(CS))/trace(CU inv(CB));

where CS is a sample covariance matrix of the features extracted from the one or more enrollment utterances;

CB is the sample covariance matrix of the features extracted from background data, and CU is the sample covariance matrix of the features extracted from the one or more test utterances.

22. The apparatus of claim 16, wherein said background model is obtained from utterances of a plurality of speakers.

23. An article of manufacture for recognizing a speaker, comprising a machine readable medium containing one or more programs which when executed implement the steps of:

extracting features from one or more enrollment utterances to generate a target speaker model based on a sample covariance matrix generated from said extracted enrollment features;

extracting features from one or more test utterances to generate a test utterance model based on a sample covariance matrix generated from said extracted test features; and

computing a sphericity ratio to recognize said speaker, said sphericity ratio comparing said test utterance model to said target speaker model and to a background model.