Speaker identification using unsupervised speech models

• US 5,946,654 A
• Filed: 02/21/1997
• Issued: 08/31/1999
• Est. Priority Date: 02/21/1997
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Dragon Systems, Inc. (Microsoft Corporation)
• Inventors: Ito, Yoshiko, Gillick, Laurence S., Newman, Michael Jack
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): SAX, ROBERT L
• Application Number:
  US08/804,061
• Time in Patent Office: 921 Days
• Field of Search: 704/246, 704/249, 704/250
• US Class Current: 704/246
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/14 Use of phonemic categorisation or speech recognition prior to speaker recognition or verification

Abstract

A speech model is produced for use in determining whether a speaker associated with the speech model produced an unidentified speech sample. First a sample of speech of a particular speaker is obtained. Next, the contents of the sample of speech are identified using speech recognition. Finally, a speech model associated with the particular speaker is produced using the sample of speech and the identified contents thereof. The speech model is produced without using an external mechanism to monitor the accuracy with which the contents were identified.

Claims

1. A method of producing a speech model associated with a speaker for use in determining whether the speaker produced an unidentified speech sample, the method comprising:

obtaining a sample of speech of the speaker; and

producing a speech model for subsequent use associated with the speaker by;

identifying words in the sample of speech using speech recognition, and

deriving a new speech model using the sample of speech and the identified words,

the speech model being derived without verifying the accuracy of the identified words and without controlling the content of the sample of speech.

2. The method of claim 1, wherein:

identifying the words comprises associating time segments of the sample of speech with corresponding sounds; and

deriving the speech model comprises causing the model to represent how the speaker produces sounds based on the associated time segments of the sample of speech.

3. The method of claim 2, wherein producing a speech model further comprises:

providing an unadapted speech model based on speech of a collection of speakers; and

adapting the unadapted speech model based on the speech sample.

4. The method of claim 3, wherein adapting the unadapted speech model comprises adapting portions of the unadapted speech model related to sounds included in the speech sample based on the corresponding time segments of the speech sample.

5. The method of claim 3, wherein adapting the unadapted speech model comprises adapting the unadapted speech model based on the identified words of the speech sample.

6. The method of claim 4, wherein the unadapted speech model comprises a monophone model.

7. The method of claim 6, wherein identifying the words comprises using a speech recognition vocabulary that includes triphone models.

8. The method of claim 5, wherein the unadapted speech model comprises a triphone model.

9. The method of claim 5, wherein the unadapted speech model represents phonemes.

10. The method of claim 9, wherein the unadapted speech model represents each phoneme as a sequence of nodes.

11. The method of claim 10, wherein the unadapted speech model represents each phoneme as a sequence of three nodes.

12. The method of claim 10, wherein the unadapted speech model represents each phoneme node using a collection of parameters that represent the frequency content of the node.

13. The method of claim 5, wherein adapting the unadapted speech model based on the sample of speech and the identified words comprises using a Bayesian adaptation approach.

14. The method of claim 1, wherein the speaker comprises a particular speaker, the method further comprising using a sample of speech of a second, unidentified speaker and the speech model of the particular speaker to determine if the particular speaker is likely to be the unidentified speaker.

15. The method of claim 1, wherein the speaker comprises a particular speaker, the method further comprising:

obtaining a second sample of speech from an unidentified speaker; and

comparing the second sample of speech to the speech model for the particular speaker to determine whether the particular speaker is likely to be the unidentified speaker.

16. The method of claim 15, wherein comparing the second sample of speech to the speech model for the particular speaker comprises generating a comparison score for the speech model based on the comparison.

17. The method of claim 16, further comprising comparing the second sample of speech to an unadapted speech model based on speech of a collection of speakers and generating a comparison score for the unadapted speech model.

18. The method of claim 17, further comprising generating a relative comparison score based on the difference between the comparison score for the speech model for the particular speaker and the comparison score for the unadapted speech model.

19. The method of claim 15, further comprising modifying the speech model using the second sample of speech.

20. The method of claim 15, further comprising modifying the speech model using the second sample of speech when the particular speaker is likely to be the unidentified speaker.

21. The method of claim 1, further comprising obtaining the sample of speech from a voice messaging system.

22. A method of identifying a speaker who produced an unidentified speech sample, comprising:

producing speech models for known speakers, wherein a speech model for each known speaker is produced by;

obtaining a sample of speech of the known speaker,

identifying words in the sample of speech of the known speaker using speech recognition, and

deriving the speech model using the sample of speech and the identified words without verifying the accuracy of the identified words and without controlling the content of the sample of speech;

obtaining a sample of speech from an unidentified speaker;

identifying words in the sample of speech from the unidentified speaker; and

comparing, based on the identified words, the sample of speech from the unidentified speaker to the speech models for the known speakers to determine whether a known speaker is likely to be the unidentified speaker.

23. The method of claim 22, further comprising indicating that none of the known speakers is likely to be the unidentified speaker when none of the speech models for the known speakers matches the sample of speech for the unidentified speaker to within a predetermined threshold.

24. The method of claim 22, further comprising identifying the unidentified speaker as the known speaker corresponding to the speech model that most closely matches the sample of speech for the unidentified speaker.

25. The method of claim 22, wherein the step of producing a speech model further comprises:

providing an unadapted speech model based on speech of a collection of speakers; and

adapting the unadapted speech model based on the speech sample for the known speaker.

26. A method of processing voice messages in a voice messaging system, comprising:

receiving a first voice message;

using speech recognition to identify words in the first voice message without controlling the content of the first voice message;

deriving a speech model for the speaker of the first voice message using the identified words; and

using the derived speech model to determine if the speaker of the first voice message is the speaker of a second voice message.

27. The method of claim 26, wherein receiving a voice message comprises receiving a voice message transmitted via a telephone line.

28. The method of claim 26, wherein the speaker does not receive an indication that a voice message will be processed to identify the speaker.

29. The method of claim 26, further comprising receiving identification of the speaker of the first voice message from a listener of the first voice message.