Discriminative gaussian mixture models for speaker verification

• US 6,411,930 B1
• Filed: 02/12/1999
• Issued: 06/25/2002
• Est. Priority Date: 11/18/1998
• Status: Expired due to Fees

• Current Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Original Assignee: Lucent Technologies, Inc. (Nokia Corporation)
• Inventors: Burges, Christopher John
• Primary Examiner(s): Chawan, Vijay B
• Application Number:
  US09/249,414
• Time in Patent Office: 1,229 Days
• Field of Search: 704/231, 704/243, 704/256, 704/244-246, 704/236, 704/238, 704/200, 704/248-251, 704/254, 704/255, 704/257, 704/240
• US Class Current: 704/240
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/06 Decision making techniques; Pattern matching strategies

Abstract

Speaker identification is performed using a single Gaussian mixture model (GMM) for multiple speakers—referred to herein as a Discriminative Gaussian mixture model (DGMM). A likelihood sum of the single GMM is factored into two parts, one of which depends only on the Gaussian mixture model, and the other of which is a discriminative term. The discriminative term allows for the use of a binary classifier, such as a support vector machine (SVM). In one embodiment of the invention, a voice messaging system incorporates a DGMM to identify the speaker who generated a message, if that speaker is a member of a chosen list of target speakers, or to identify the speaker as a “non-target” otherwise.

Claims

1. In a method for performing speaker verification, wherein the identified speaker is determined to belong to a speaker set comprising N target speas, where N>

1, the improvement comprising;

adding discrimtion to a Gaussian mixture model (GMM);

wherein said discrination is based at least upon characteristics of all of said speaks; and

wherein a single GMM is used for the speaker set.

2. The method of claim 1 further comprising the step of using binary classifiers for selecting the identified speaker.

3. The method of claim 2 wherein the binary classifiers are support vector machines.

4. The method of claim 1 further comprising the steps of:

training the discriminative GMM on speech;

generating support vector machine training data by clustering data resulting from the GMM training step; and

using at least one support vector machine for selecting the identified speaker.

5. The method of claim 4 wherein the speech used for GMM training includes speech from at least one target speaker.

6. The method of claim 4, wherein the using step uses a support vector machine to process a speech sample if other processing of the speech sample relative to a predefined threshold indicates that the speech sample exceeds the predefined threshold.

7. The method of claim 6, wherein the predefined threshold is a measure of a likelihood that the speech sample came from a GMM cluster.

8. The method of claim 1 further comprising the steps of:

training the GMM on speech;

clustering data resulting from the training step;

associating N support vector machines for each cluster; and

using at least one of the support vector machines for selecting the identified speaker.

9. The method of claim 8, wherein the speech used for GMM training includes speech from at least one target speaker.

10. The method of claim 8, wherein the using step uses a support vector machine to process a speech sample if other processing of the speech sample relative to a predefined threshold indicates that the speech sample exceeds the predefined threshold.

11. The method of claim 10, wherein the predefined threshold is a measure of a likelihood that the speech sample came from a cluster of GMM data.

12. A method for use in a voice messaging system for verfying a speaker, the method comprising the steps of:

training a Gausian mixture model (GMM) on speech, wherein the GMM is used for N speakrs, where N>

1 and wherein said training is based at least upon characteristics of all of said N speakers;

generatmg support vector machine training data by clustering data resulting from the training step; and

using at least one support vector machine to determine that the speaker belongs to the set of said N speakers.

13. The method of claim 12 wherein the speech used for GMM training includes speech from at least one of the N speakers.

14. The method of claim 12, wherein the using step uses a support vector machine to process a speech sample if other processing of the speech sample relative to a predefined threshold indicates that the speech sample exceeds the predefined threshold.

15. The method of claim 14, wherein the predefined threshold is a measure of a likelihood that the speech sample came from a cluster of GMM data.

16. The method of claim 12, wherein the voice messaging system comprises a server and at least one client and the steps of the method of claim 12 are performed in either one.

17. The method of claim 12, wherein the voice messaging system comprises a server and at least one client and the steps of the method of claim 12 are distributed between them.

18. A voice messaging system comprising:

a client;

a server for (a) storing a voice message, (b) verifying a speaker of the voice message as being from a set of N target speakers, where N>

1, and (c) providing speaker verification information to the client;

wherein the server uses a discriminative Gaussian mixtue model (GMM) for verifying the speaker, and wherein the discriminative GMM is based at least upon characteristics of all of the N target speakers; and

wherein the speaker is identified as being one of said N speakers.

19. The system of claim 18 wherein the speaker verification information is further incorporated into a confidence level.

20. Voice messaging apparatus comprising:

means for receiving a voice message; and

a processor that verifies a speaker of the voice message as belonging to a set of N possible speakers, where N>

1;

wherein the processor uses a discriminative Gaussian Mixture Model (GMM) for verifying the speaker, and where the same discriminative GMM is used for all the N possible speakers and wherein the discriminative GMM is based at least upon characteristics of all N possible speakers.

21. An article of manirware for use in performing speaker verification comprising a machine readable medium comprising one or more programs which when executed implement the step of:

using a discriminative Gaussian mixture model (DGMM) for use in associating a message with an unspecified speaker from a set of N possible speakers, where N>

1, such that the DGMM uses one GMM for the N possible speakers;

wherein said DGMM is based at least upon characteristics of all of the N possible speakers.

22. The program of claim 21 further comprising the step of using a binary classifier for selecting the unspecified speaker.

23. The program of claim 22 wherein the binary classifier is a support vector machine.

24. The program of claim 21 further comprising the steps of:

training the DGMM on speech;

generating support vector machine training data by clustering data resulting from the training step; and

using at least one support vector machine for selecting the associated speaker.

25. The program of claim 24 wherein the speech used for DGMM training includes speech from the N possible speakers.

26. The program of claim 24, wherein the using step uses one or more support vector machines to process a speech sample if other processing of the speech sample relative to a predefined threshold indicates that the speech sample exceeds the predefined threshold.

27. The program of claim 26, wherein the predefined threshold is a measure of a likelihood that the speech sample came from one of the clusters.