Providing a confidence measure for speaker diarization

1. A computer implemented method for providing a confidence measure for speaker diarization, comprising:

receiving an audio session as unsegmented audio data;

computing a spectral ratio of principal component analysis (PCA) of sections of the received audio session by a ratio between the largest eigenvalue and the second largest eigenvalue;

using the PCA spectral ratio as a confidence measure for speaker diarization processing.

2. The method as claimed in claim 1 , wherein computing a spectral ratio of principal component analysis includes:

dividing the session into sections;

parameterizing each section by a time dependent probability density function to define a supervector space;

creating a basis for the intra-speaker variability using principal component analysis (PCA);

projecting all the sections according to the basis to provide compensated vectors;

estimating the inter-speaker variability of the sections in projected space;

creating a basis for the inter-speaker variability on the compensated vectors using PCA.

3. The method as claimed in claim 1 , wherein using the PCA spectral ratio as a confidence measure for speaker diarization includes:

determining if the spectral ratio is a large value indicating that the separation of two speakers is good; and

determining if the spectral ratio is a low value indicating that the separation is not good.

4. The method as claimed in claim 1 , including:

using the confidence measure for filtering out bad segmentations in two-speaker diarization.

5. The method as claimed in claim 1 , including:

using the confidence measure for guiding online diarization.

6. The method as claimed in claim 1 , including:

using the confidence measure to classify between single speaker and two speaker utterances.

7. The method as claimed in claim 2 , wherein parameterizing each section by a time dependent probability density function to define a supervector space uses a Gaussian Mixture Model.

8. The method as claimed in claim 2 , wherein creating a basis for the intra-speaker variability using PCA includes:

computing the difference between each two consecutive supervectors;

estimating a scatter measure of the difference vectors;

applying PCA to the scatter measure and retaining the first n eigenvalues.

9. The method as claimed in claim 8 , wherein projecting all the sections according to the basis includes:

compensating each original supervector by removing the projection of each original supervector on selected n eigenvectors of the PCA.

10. The method as claimed in claim 2 , wherein estimating the inter-speaker variability of the sections in projected space, includes:

computing a scatter measure of the compensated supervectors; and

applying PCA to the scatter measure.

11. The method as claimed in claim 2 , wherein dividing a speech session into sections divides the speech session into evenly spaced overlapping sections.

12. A computer program product for providing a confidence measure for speaker diarization, the computer program product comprising:

a non-transitory computer readable storage medium having computer readable program code

embodied therewith, the computer readable program code comprising:

computer readable program code configured to:

receive an audio session as unsegmented audio data;

compute a spectral ratio of principal component analysis (PCA) of sections of the received audio session by a ratio between the largest eigenvalue and the second largest eigenvalue;

use the PCA spectral ratio as a confidence measure for speaker diarization processing.

13. A system for providing a confidence measure for speaker diarization, comprising:

a processor;

a session receiver and dividing component for receiving an audio session as unsegmented audio data and dividing it into sections;

a spectral ratio component for computing a spectral ratio of principal component analysis (PCA) of the sections of the received audio session by a ratio between the largest eigenvalue and the second largest eigenvalue;

a confidence measure applying component for using the PCA spectral ratio as a confidence measure for speaker diarization processing.

14. The system as claimed in claim 13 , including a confidence measure component including:

a parameterizing component for parameterizing each section by a time dependent probability density function to define a supervector space;

an intra-speaker variability component for creating a basis for the intra-speaker variability using principal component analysis (PCA);

a section projecting component for projecting all the sections according to the basis to provide compensated vectors;

an inter-speaker variability estimating component for estimating the inter-speaker variability of the sections in projected space for creating a basis for the inter-speaker variability on the compensated vectors using PCA.

15. The system as claimed in claim 13 , wherein a confidence measure applying component for using the PCA spectral ratio as a confidence measure for speaker diarization processing includes:

determining if the spectral ratio is a large value indicating that the separation of two speakers is good; and

determining if the spectral ratio is a low value indicating that the separation is not good.

16. The system as claimed in claim 13 , wherein a confidence measure applying component for using the PCA spectral ratio as a confidence measure for speaker diarization processing includes:

a bad segment filtering component for using the confidence measure for filtering out bad segmentations in two-speaker diarization.

17. The system as claimed in claim 13 , wherein a confidence measure applying component for using the PCA spectral ratio as a confidence measure for speaker diarization processing includes:

an online component for using the confidence measure for guiding online diarization.

18. The system as claimed in claim 13 , wherein a confidence measure applying component for using the PCA spectral ratio as a confidence measure for speaker diarization processing includes:

a number of speakers component for using the confidence measure to classify between single speaker and two speaker utterances.

19. The system as claimed in claim 14 , wherein the parameterizing component parameterizes each section using a Gaussian Mixture Model.

20. The system as claimed in claim 14 , wherein the intra-speaker variability component is for:

computing the difference between each two consecutive supervectors;

estimating a scatter measure of the difference vectors;

applying PCA to the scatter measure and retaining the first n eigenvalues.

21. The system as claimed in claim 14 , wherein the section projecting component is for:

compensating each original supervector by removing the projection of each original supervector on selected n eigenvectors of the PCA.

22. The system as claimed in claim 14 , wherein the inter-speaker variability estimating component is for:

computing a scatter measure of the compensated supervectors; and

applying PCA to the scatter measure.

23. The system as claimed in claim 13 , wherein the session receiver and dividing component divides the speech session into evenly spaced overlapping sections.