Biometric speaker identification

• US 8,719,018 B2
• Filed: 10/25/2010
• Issued: 05/06/2014
• Est. Priority Date: 10/25/2010
• Status: Active Grant

• Current Assignee: Leidos Innovations Technology, Inc.
• Original Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Inventors: Dinerstein, Jonathan J.
• Primary Examiner(s): MCFADDEN, SUSAN IRIS
• Application Number:
  US12/911,140
• Publication Number:
  US 20120101822A1
• Time in Patent Office: 1,289 Days
• Field of Search: 704/246
• US Class Current: 704/246
• CPC Class Codes: G10L 17/04 Training, enrolment or model building

Abstract

A biometric speaker-identification apparatus is disclosed that generates ordered speaker-identity candidates for a probe based on prototypes. Probe match scores are clustered, and templates that correspond to clusters having top M probe match scores are compared with the prototypes to obtain template-prototype match scores. The probe is also compared with the prototypes, and those templates corresponding to template-prototype match scores that are nearest to probe-prototype match scores are selected as speaker-identity candidates. The speaker-identity candidates are ordered based on their similarity to the probe.

Claims

1. A biometric speaker-identification apparatus that generates one or more speaker-identity candidates for a speaker based on probe match scores obtained by performing a voice matching operation between a probe and templates in a biometric corpus, comprising:

a plurality of prototypes; and

a speaker-identification processor coupled to the biometric corpus, the speaker-identification processor configured to select templates of the biometric corpus as the speaker-identity candidates based on the prototypes,

wherein the speaker-identification processor is further configured to;

perform the voice matching operation between the speaker-identity candidates and the templates in the biometric corpus to obtain one or more speaker-identity-candidate match scores;

perform a similarity measurement between the speaker-identity-candidate match scores and the probe match scores to obtain one or more similarity values; and

order the speaker-identity candidates based on the similarity values, and

wherein the similarity measurement for an i^th speaker-identity candidate is a dot product defined as;

DOT_i=SUM(tms_it*pms_t),

where the sum is taken over all the templates in the biometric corpus, tms_it is a template match score between the i^th speaker-identity candidate and a t^th template, and pms_t is a probe match score for the t^th template.

2. The apparatus of claim 1 wherein the speaker-identification processor is further configured to:

perform the voice matching operation between the probe and the prototypes to obtain probe-prototype match scores;

perform the voice matching operation between first templates selected from the biometric corpus and the prototypes to obtain template-prototype match scores; and

select as the speaker-identity candidates one or more second templates corresponding to template-prototype match scores that are nearest to the probe-prototype match scores based on a nearness measurement.

3. The apparatus of claim 2 wherein the speaker-identification processor is further configured to:

cluster the probe match scores until probe match scores in each of K clusters have a deviation less than about ten times a deviation in all the probe match scores; and

select templates that belong to clusters that have probe match scores greater than M as the first templates.

4. The apparatus of claim 2 wherein the speaker-identification processor is further configured to:

cluster the probe match scores into K clusters; and

select templates from the biometric corpus corresponding to ones of the K clusters having probe match scores that exceed M as the first templates.

5. The apparatus of claim 4 wherein the speaker-identification processor is configured to start clustering the probe match scores with K=1 and repeats the clustering with K incremented by 1 until a difference between a maximum probe match score and a minimum probe match score in each of the K clusters is less than about ten times a difference between a maximum probe match score and a minimum probe match score in all the probe match scores.

6. The apparatus of claim 2 wherein the nearness measurement is a Euclidian distance in a P dimensional hyperspace spanned by the prototypes, the template-prototype match scores and the probe-prototype match scores are coordinates of points in the P dimensional hyperspace, and the speaker-identification processor is further configured to select as the speaker-identity candidates templates corresponding to template-prototype match scores that are within an R radius of a probe point corresponding to the probe-prototype match scores.

7. A biometric speaker-identification apparatus that generates a speaker-identity candidate for a speaker based on probe match scores obtained by performing a voice matching operation between a probe and templates in a biometric corpus, comprising:

a plurality of prototypes;

a speaker-identification processor coupled to the biometric corpus, the speaker-identification processor configured to;

group the probe match scores into K clusters;

select templates corresponding to ones of the K clusters having probe match scores that exceed M;

perform the voice matching operation between the selected templates and the prototypes to obtain template-prototype match scores;

perform the voice matching operation between the probe and the prototypes to obtain probe-prototype match scores;

select as speaker-identity candidates one or more templates corresponding to template-prototype match scores that are nearest to the probe-prototype match scores based on a nearness measurement

perform the voice matching operation between the speaker-identity candidates and the templates in the biometric corpus to obtain one or more speaker-identity-candidate match scores;

perform a similarity measurement between the speaker-identity-candidate match scores and the probe match scores to obtain one or more similarity values; and

order the speaker-identity candidates based on the similarity values,

wherein the similarity measurement for an i^th speaker-identity candidate is a dot product defined as;

DOT_i=SUM(tms_it*pms_t),

where the sum is taken over all the templates in the biometric corpus, tms_it is a template match score between the i^th speaker-identity candidate and a t^th template, and pms_t is a probe match score for the t^th template.

8. The apparatus of claim 7 wherein the speaker-identification processor is further configured to:

perform the voice matching operation between the speaker-identity candidates and the templates in the biometric corpus to obtain speaker-identity-candidate match scores;

perform a similarity measurement between the speaker-identity-candidate match scores and the probe match scores to obtain similarity values; and

order the speaker-identity candidates based on the similarity values.

9. A biometric speaker-identification method performed by processing circuitry, the method comprising:

storing a plurality of prototypes in a memory;

selecting templates of a biometric corpus as one or more speaker-identity candidates for a probe based on the prototypes;

performing a voice matching operation between the speaker-identity candidates and the templates in the biometric corpus to obtain speaker-identity-candidate match scores;

performing a similarity measurement between the speaker-identity-candidate match scores and probe match scores to obtain similarity values; and

ordering the speaker-identity candidates based on the similarity values,

wherein the similarity measurement for an i^th speaker-identity candidate is a dot product defined as;

DOT_i=SUM(tms_it*pms_t),

where the sum is taken over all the templates in the biometric corpus, tms_it is a template match score between the i^th speaker-identity candidate and a t^th template, and pms_t is a probe match score for the t^th template.

10. The method of claim 9 further comprising:

performing the voice matching operation between the probe and the prototypes to obtain probe-prototype match scores;

selecting first templates from the biometric corpus;

performing the voice matching operation between the first templates and the prototypes to obtain template-prototype match scores; and

selecting as the speaker-identity candidates one or more second templates corresponding to template-prototype match scores that are nearest to the probe-prototype match scores based on a nearness measurement.

11. The method of claim 10 furthering comprising:

clustering the probe match scores until probe match scores in each of K clusters have a deviation less than about ten times a deviation in the probe match scores; and

selecting templates that belong to clusters that have probe match scores greater than M as the first templates.

12. The method of claim 10 further comprising:

storing, in the memory, probe match scores obtained by performing a voice matching operation between the probe and the templates of the biometric corpus;

clustering the probe match scores into K clusters; and

selecting templates corresponding to ones of the K clusters having probe match scores that exceed M as the first templates.

13. The method of claim 12 wherein the clustering comprises:

setting K=1;

performing a clustering process; and

repeating the clustering process with K incremented by 1 each time until in each of the K clusters a difference between a maximum probe match score and a minimum probe match score is less than about ten times a difference between a maximum probe match score and a minimum probe match score in all the probe match scores.

14. The method of claim 11 wherein the nearness measurement is a Euclidian distance, the method further comprising:

spanning a P dimensional hyperspace having the prototypes as axes; and

selecting as the speaker-identity candidates templates corresponding to template-prototype match scores that are within an R radius of a probe point corresponding to the probe-prototype match scores, wherein template-prototype match scores and probe-prototype match scores are coordinates of points in the P dimensional hyperspace.

15. The method of claim 9 further comprising:

grouping probe match scores into K clusters;

selecting templates corresponding to ones of the K clusters having probe match scores that exceed M;

performing the voice matching operation between the selected templates and the prototypes to obtain template-prototype match scores;

performing the voice matching operation between the probe and the prototypes to obtain probe-prototype match scores; and

selecting as the speaker-identity candidates templates corresponding to template-prototype match scores that are nearest to the probe-prototype match scores based on a Euclidian distance measurement in a P dimensional space spanned by the prototypes.

16. The method of claim 15 further comprising:

performing the voice matching operation between the speaker-identity candidates and the templates in the biometric corpus to obtain speaker-identity-candidate match scores;

performing a dot product between the speaker-identity-candidate match scores and the probe match scores to obtain similarity values; and

ordering the speaker-identity candidates based on the similarity values.