Method and apparatus for speaker spotting

• US 8,078,463 B2
• Filed: 11/23/2004
• Issued: 12/13/2011
• Est. Priority Date: 11/23/2004
• Status: Active Grant

• Current Assignee: Nice Ltd
• Original Assignee: Nice Systems Limited (Nice Ltd)
• Inventors: Wasserblat, Moshe, Zigel, Yaniv, Pereg, Oren
• Primary Examiner(s): Vo, Huyen X.
• Application Number:
  US10/996,811
• Publication Number:
  US 20060111904A1
• Time in Patent Office: 2,576 Days
• Field of Search: 704/246, 704/275, 704/245, 704/250, 704/243, 704/239, 704/214, 704/247, 704/270.1, 704/249, 704/238, 704/240, 704/244, 382/224, 455/312
• 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 computerized method for spotting an at least one call interaction out of a multiplicity of call interactions, in which an at least one target speaker participates, the method comprising:

capturing at least one target speaker speech sample of the at least one target speaker by a speech capture device;

generating by a computerized engine a multiplicity of speaker models based on a multiplicity of speaker speech samples from the at least one call interaction;

matching by a computerized server the at least one target speaker speech sample with speaker models the multiplicity of speaker models to determine a target speaker model;

determining a score for each call interaction of the multiplicity of call interactions according to a comparison between the target speaker model and the multiplicity of speaker models; and

based on scores that are higher than a predetermined threshold, determining call interactions, of the multiplicity of call interactions, in which the at least one target speaker participates.

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 target speaker speech sample;

and

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

4. The method for claim 1 wherein the step of searching further comprises pre-processing the at least one target speaker speech sample;

and

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

5. The method according to claim 3 or claim 4, wherein the pre-processing step comprises segmenting the at least one call interaction into speech frames;

and eliminating an at least one frame including noise or silence.

6. The method of claim 1 wherein the step of generating further comprises estimating the multiplicity of speaker models based on an at least one extracted feature vector.

7. The method of claim 1 wherein the step of generating further comprises storing the multiplicity of speaker models with additional speaker data in a speaker model database.

8. The method of claim 7 wherein the additional speaker data includes an at least one pointer to the at least one target speaker speech sample or to a portion of the at least one target speaker speech sample.

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

10. The method of claim 9 further comprises sorting the probabilistic scores and obtaining a score threshold value for the matching of the at least one target speaker speech sample with a speaker model of the multiplicity of speaker models.

11. The method of claim 10 further comprising indicating speaker speech models of the multiplicity of speaker models with scores above the score threshold value.

12. The method of claim 1 wherein the step of searching further comprises inserting the at least one target speaker speech sample into a sorted calls data structure.

13. The method of claim 1 wherein the step of searching comprises a step of fast searching of the multiplicity of speaker models in order to eliminate a portion of the multiplicity of speaker models.

14. The method of claim 1 wherein the step of searching further comprises testing a quality parameter of an at least one frame within the multiplicity of call interactions or within the multiplicity of speaker speech samples, the at least one frame containing an at least one feature vector, and eliminating the at least one frame if the quality is below a predetermined threshold.

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

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

17. The method of claim 1 wherein the spotting of the at least one target speaker is performed online.

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

19. The method of claim 1 further comprises disguising the at least one target speaker by distorting a 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 at least one target speaker and extracting information from the interactions.

21. The method of claim 20 further comprises displaying the historical speech-based interactions associated with the at least one target speaker and displaying with the extracted information.

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

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

segmenting an at least one speaker from the summed target speakers of at least one call interaction using only audio information of the at least one call interaction; and

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

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

24. The method of claim 22 wherein the separating of the at least one speaker from the summed target speakers in the at least one call interaction is performed manually.

25. The method of claim 1, further comprising the step of outputting at least one of the multiplicity of call interactions in which the at least one target speaker participates.

26. The method of claim 1 wherein matching the at least one target speaker speech sample with the multiplicity of speaker models comprises comparing feature vector values, extracted from the at least one target speaker speech sample, with the multiplicity of speaker models.

27. The method of claim 1, wherein matching the at least one target speaker speech sample with the multiplicity of speaker models further comprises utilizing additional information linked to the multiplicity of speaker models.

28. The method of claim 27, wherein the additional information comprises at least one of a speaker profile, case data, phone equipment data, target data, call data, a call location, or warrant data, or any combination thereof.

29. A computerized apparatus for spotting an at least one call interaction out of a multiplicity of call interactions in which a target speaker participates, the apparatus comprising:

a training computerized component configured for generating a multiplicity of speaker models based on a multiplicity of speaker speech samples from the at least one call interaction; and

a speaker spotting computerized component configured for matching the target speaker speech sample with speaker models of the multiplicity of speaker models to determine a target speaker model,

determining a score for each call interaction of the multiplicity of call interactions according to a comparison between the target speaker model and the multiplicity of speaker models, and

based on scores that are higher than a predetermined threshold, determining call interactions, of the multiplicity of call interactions, and in which the at least one target speaker participates.

30. The apparatus of claim 29 further comprises a speaker model storage component to store the multiplicity of speaker models based on the multiplicity of speaker speech samples.

31. The apparatus of claim 30 wherein the speaker model storage component comprises a speaker model database to hold the multiplicity of speaker models.

32. The apparatus of claim 29 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 vector;

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

33. The apparatus of claim 29 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 at least one speaker model associated with the speaker speech.

34. The apparatus of claim 33 wherein the speaker model database comprises at least one speaker models and at least one feature probability density function parameter 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.

35. The apparatus of claim 29 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 target speaker speech sample.

36. The apparatus of claim 29 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 multiplicity of speaker models.

37. The apparatus of claim 29 further comprising a call logging system for capturing the target speaker speech sample or at least one of the multiplicity of call interactions.

38. The apparatus of claim 29 further comprising a calls database for storing at least one of the multiplicity of call interactions.