Method and apparatus for speaker spotting

• US 20060111904A1
• Filed: 11/23/2004
• Published: 05/25/2006
• Est. Priority Date: 11/23/2004
• Status: Active Grant

• Current Assignee: Nice Ltd
• Original Assignee: Nice Systems Limited (Nice Ltd)
• Inventors: Wasserblat, Moshe, Pereg, Oren, Zigel, Yaniv
• Granted Patent: US 8,078,463 B2
• Time in Patent Office: Days
• US Class Current: 704/246
• CPC Class Codes: G10L 17/00 Speaker identification or verification techniques

Abstract

A method and apparatus for spotting a target speaker within a call interaction by generating speaker models based on one or more speaker'"'"'s speech; and by searching for speaker models associated with one or more target speaker speech files.

Claims

1. A method for spotting an at least one target speaker within at least one call interaction, the method comprising:

generating from the at least one call interaction an at least one speaker model of the at least one speaker based on an at least one speaker speech sample; and

searching for the at least one target speaker using the at least one speech sample of the target speaker and the at least one speaker model.

2. The method of claim 1 wherein the step of generating comprises obtaining the at least one speaker speech sample from a multi-speaker speech database.

3. The method of claim 1 wherein the step of generating further comprises

pre-processing the at least one speaker speech sample;

and

extracting an at least one feature vector from the at least one speaker speech sample.

4. The method of claim 1 wherein the step of generating further comprises

estimating the at least one speaker model based on an at least one extracted feature vector.

5. The method of claim 1 wherein the step of generating further comprises

storing the at least one speaker model with additional speaker data in a speaker model database.

6. The method of claim 5 wherein the additional speaker data includes an at least one pointer to the speaker speech sample or to a portion of the speaker speech sample.

7. The method of claim 1 wherein the step of searching comprises obtaining the target speaker speech sample from a speech capture device or from a pre-stored speech recording.

8. The method for claim 1 wherein the step of searching further comprises

pre-processing the target speaker speech sample;

and

extracting an at least one feature vector from the target speaker speech sample.

9. The method of claim 1 wherein the step of searching further comprises

calculating probabilistic scores, indicating the matching of the target speaker speech sample with the at least one speaker model.

10. The method of claim 1 wherein the step of searching further comprises

inserting the at least one target speaker speech into a sorted calls data structure.

11. The method of claim 1 wherein the step of searching comprises fast searching of the at least one speaker speech model.

12. The method of claim 1 wherein the step of searching further comprises

testing the quality of an at least one frame containing an at least one feature vector.

13. The method of claim 9 further comprises sorting the probabilistic scores and obtaining a score threshold value for the speaker speech models.

14. The method of claim 13 further comprising indicating speaker speech models with scores above the threshold.

15. The method of claim 1 further comprises extracting speech characteristics of a speech sample of a speaker associated with a target speaker.

16. The method of claim 1 wherein the spotting of the target speaker is performed offline.

17. The method of claim 16 wherein the spotting of the target speaker is performed online.

18. The method of claim 1 further comprises masking speech based interactions associated with an at least one target speaker.

19. The method of claim 1 further comprises disguising the identity of the at least one target speaker by distorting the speech pattern of the at least one target speaker.

20. The method of claim 1 further comprises finding historical speech-based interactions associated with the target speaker and extracting information from the interactions.

21. The method of claim 1 further comprises displaying the historical speech-based interactions associated with the target speaker and displaying with the extracted information.

22. A method of detecting an at least one speaker within at least one call interaction, the method comprising:

determining the number of summed speakers from the at least one call interaction;

segmenting the speakers from the summed target speakers of at least one call interaction;

separating the speakers from the summed target speakers of the at least one call interaction.

23. The method of claim 22 wherein the separation of the target speakers from the summed target speakers in the at least one call interaction is performed automatically.

24. The method of claim 23 wherein the separation of the target speakers from the summed target speakers in the at least one call interaction is performed manually.

25. An apparatus for spotting a target speaker, the apparatus comprising a training component to generate at least one speaker model based on at least one speaker speech;

and a speaker spotting component to match the at least one speaker model to an at least one target speaker speech.

26. The apparatus of claim 25 further comprises a speaker model storage component to store the at least one speaker model based on the at least one speaker speech.

27. The apparatus of claim 25 wherein the training component comprises a speaker speech pre-processor module to pre-process an at least one speaker speech sample and an at least one speech feature vectors;

and

an extraction module to extract the at least one speech feature vectors from the pre-processed at least one speaker speech sample.

28. The apparatus of claim 25 wherein the training component comprise a speaker model estimation module to generate at least one reference speaker model associated with an extracted at least one speech feature vector;

and a speaker models database to store generated at least one speaker model associated with the speaker speech.

29. The apparatus of claim 28 wherein the speaker model database comprises at least one speaker models and at least one feature probability density function parameters associated with the at least one speaker speech;

the at least one speaker speech sample associated with the at least one speaker model; and

additional speaker information for storing speaker data.

30. The apparatus of claim 26 wherein the speaker model storage component comprises a speaker model database to hold the at least one speaker models.

31. The apparatus of claim 25 wherein the speaker spotting component further comprises a target speaker speech feature vectors extraction module to extract at least one target speaker speech feature vectors from the pre-processed target speaker speech sample.

32. The apparatus of claim 25 wherein the speaker spotting component further comprises a score calculation component to score target speaker speech and to match the target speaker speech to the at least one speaker models.