Efficient recursive clustering based on a splitting function derived from successive eigen-decompositions

1. An apparatus for enrollment of a speaker in a speaker verification system comprising:

an arrangement for obtaining speech and audio data as input data; and

an arrangement for clustering of said speech and audio data, said arrangement for clustering further comprising:

an arrangement for creating a predetermined number of non-overlapping subsets of the input data;

said arrangement for creating a predetermined number of non-overlapping subsets by splitting the input data recursively, wherein said arrangement for creating a predetermined number of non-overlapping subsets determines an eigenvector decomposition relating to the input data;

said clustering being independent of any model wherein the splitting of the input data into a predetermined number of non-overlapping subsets occurs independent of a model;

wherein there is no variability in the clustering due to randomness.

2. The apparatus according to claim 1 , wherein said arrangement for creating a predetermined number of non-overlapping subsets splits the input data into at least two sets of output data.

3. The apparatus according to claim 2 , wherein said arrangement for creating a predetermined number of non-overlapping subsets:

splits the at least two sets of output data recursively; and

repeats the recursive splitting of output data sets until the predetermined number of non-overlapping subsets is obtained.

4. The apparatus according to claim 2 , wherein said arrangement for creating a predetermined number of non-overlapping subsets determines a vector of projection coefficients onto the set of eigenvectors in the eigenvector decomposition.

5. The apparatus according to claim 4 , wherein said arrangement for creating a predetermined number of non-overlapping subsets determines a probability distribution relating to the vector of projection coefficients.

6. The apparatus according to claim 5 , wherein said arrangement for creating a predetermined number of non-overlapping subsets:

assigns at least one threshold relating to the probability distribution; and

yields the at least two sets of output data based on the relation to the threshold of a value associated with a function relating to the projection coefficients.

7. The apparatus according to claim 6 , wherein there are N−1 thresholds, where N is the number of sets of output data to be yielded.

8. The apparatus according to claim 7 , wherein each threshold is a value of the function relating to the projection coefficients for which the probability distribution equals m/N, where m is a number from 1 to N−1.

9. The apparatus according to claim 1 , further comprising:

an arrangement for, after the splitting of the input data is complete, building a statistical model for pattern classification for each of the classes using any desired technique.

10. A method for enrollment of a speaker in a speaker verification system comprising:

obtaining speech and audio data as input data; and

clustering of said speech and audio data, said clustering further comprising:

creating a predetermined number of non-overlapping subsets of the input data;

said step of creating a predetermined number of non-overlapping subsets comprising splitting the input data recursively, wherein said splitting step comprises determining an eigenvector decomposition relating to the input data;

said clustering being independent of any model wherein the splitting of the input data into a predetermined number of non-overlapping subsets occurs independent of a model;

wherein there is no variability in the clustering due to randomness.

11. The method according to claim 10 , wherein said splitting step comprises initially splitting the input data into at least two sets of output data.

12. The method according to claim 11 , wherein said splitting step comprises:

splitting the at least two sets of output data recursively; and

repeating the recursive splitting of output data sets until the predetermined number of non-overlapping subsets is obtained.

13. The method according to claim 11 , wherein said splitting step further comprises determining a vector of projection coefficients onto the set of eigenvectors in the eigenvector decomposition.

14. The method according to claim 13 , wherein said splitting step further comprises determining a probability distribution relating to the vector of projection coefficients.

15. The method according to claim 14 , wherein said splitting step further comprises:

assigning at least one threshold relating to the probability distribution; and

yielding the at least two sets of output data based on the relation to the threshold of a value associated with a function relating to the projection coefficients.

16. The method according to claim 15 , wherein there are N−1 thresholds, where N is the number of sets of output data to be yielded.

17. The method according to claim 16 , wherein each threshold is a value of the function relating to the projection coefficients for which the probability distribution equals m/N, where m is a number from 1 to N−1.

18. The method according to claim 10 , wherein, after splitting of the input data is complete, building a statistical model for pattern classification for each of the classes using any desired technique.

19. A program storage device readable by machine, for enrollment of a speaker in a speaker verification system tangibly embodying a program of instructions executable by the machine to perform method steps for facilitating clustering of speech and audio data, said method comprising the steps of:

obtaining speech and audio data as input data; and

clustering of said speech and audio data, said clustering further comprising:

creating a predetermined number of non-overlapping subsets of the input data;

said step of creating a predetermined number of non-overlapping subsets comprising splitting the input data recursively, wherein said splitting step comprises determining an eigenvector decomposition relating to the input data;

said clustering being independent of any model wherein the splitting of the input data into a predetermined number of non-overlapping subsets occurs independent of a model;

wherein there is no variability in the clustering due to randomness.