Background learning of speaker voices

• US 20030088414A1
• Filed: 05/07/2002
• Published: 05/08/2003
• Est. Priority Date: 05/10/2001
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Koninklijke Philips Electronics N.V. (Koninklijke Philips N.V.)
• Inventors: Cheng, Jyh-Min, Chu, Ya-Cherng, Huang, Chao-Shih, Tsai, Wei-Ho
• Granted Patent: US 7,171,360 B2
• Time in Patent Office: Days
• US Class Current: 704/246
• CPC Class Codes: G10L 15/07 to the speaker
• CPC Class Codes: G10L 17/04 Training, enrolment or model building

Abstract

A speaker identification system includes a speaker model generator 110 for generating a plurality of speaker models. To this end, the generator records training utterances from a plurality of speakers in the background, without prior knowledge of the speakers who spoke the utterances. The generator performs a blind clustering of the training utterances based on a predetermined criterion. For each of the clusters a corresponding speaker model is trained.

A speaker identifier 130 identifies a speaker determining a most likely one of the speaker models for an utterance received from the speaker. The speaker associated with the most likely speaker model is identified as the speaker of the test utterance.

Claims

1. A method of automatically identifying a speaker;

the method including;

identifying a speaker by;

receiving a test utterance from the speaker;

determining a most likely one of a plurality of speaker models for the test utterance; and

identifying the speaker associated with the most likely speaker model as the speaker of the test utterance;

wherein the method includes generating the plurality of speaker models in the background by;

receiving training utterances from the plurality of speakers in the background, without prior knowledge of the speakers who spoke the respective training utterances;

blind clustering of the training utterances based on a predetermined criterion; and

training for each of the clusters a corresponding speaker model;

each of the models representing a speaker.

2. A method as claimed in claim 1, wherein the step of blind clustering of the training utterances x_i, i<

N based on the predetermined criterion includes;

modeling each respective one of the training utterances x_i by a respective model λ

_i;

calculating for each training utterance x_i a corresponding likelihood vector L_i, where each vector element L_ij, 1≦

j≦

N represents a likelihood of the training utterance x_i against a respective one of the models λ

_j;

determining for each training utterance x_i a corresponding ranking vector F_i, where each element F_ij of the ranking vector F_i is assigned a ranking value representing a ranking of the corresponding likelihood L_ij compared to the other elements of the likelihood vector L_i such that a higher likelihood value of L_ij is reflected by a higher ranking value of F_ij;

clustering the training utterances x_i based on a criterion that a minimum in a distance measure between F_i and F_j indicates that the training utterances x_i and x_j originate from a same speaker.

3. A method as claimed in claim 2, wherein the ranking is such that η

lowest likelihood values of the elements L_ij of the likelihood vector L_i are represented by distinct values of the corresponding elements F_ij of the ranking vector F_i, and that the remaining N-η

elements L_ij of the likelihood vector L_i are represented by a same predetermined ranking value of the corresponding elements F_ij of the ranking vector F_i, where η

represents an expected number of training utterances per cluster, and the predetermined ranking value being lower than any of the η

distinct ranking values.

4. A method as claimed in claim 1, wherein the method includes:

receiving an enrolment utterance from a speaker,

determining a most likely one of a plurality of speaker models for the enrolment utterance;

receiving identifying information of the speaker; and

storing the identifying information in association with the most likely speaker model.

5. A method as claimed in claim 4, wherein the method includes:

verifying whether a likelihood of the most likely speaker model is above a predetermined threshold; and

if the likelihood is below the predetermined threshold, requesting a further utterance from the speaker, and iteratively

receiving the further utterance;

adapting the most likely speaker model with the further utterance; and

determining the likelihood of the adapted speaker model;

until the likelihood is above the predetermined threshold.

6. A method as claimed in claim 1, wherein the steps of recording training utterances, blind clustering the utterances and training the speaker models is performed iteratively until a predetermined level of confidence has been achieved.

7. A method as claimed in claim 6, wherein in response to achieving the predetermined confidence level the speaker is automatically requested to provide information identifying the speaker, followed by receiving the identifying information and storing the identifying information in association with the most likely speaker model.

8. A method as claimed in claim 1, wherein the method includes, in response to having identified the speaker, automatically retrieving a personal profile for interacting with a CE device.

9. A method as claimed in claim 1, wherein the method includes recognizing the test utterance used for identifying the speaker as a voice command;

and executing the recognized voice command in a speaker-dependent way.

10. A computer program product operative to cause a processor to perform the method as claimed in claim 1.

11. A system for automatically identifying a speaker;

the system includes;

a speaker identifier operative to identify a speaker by;

receiving a test utterance from the speaker;

determining a most likely one of a plurality of speaker models for the test utterance; and

identifying the speaker associated with the most likely speaker model as the speaker of the test utterance; and

a speaker model generator operative to generate the plurality of speaker models,

wherein the speaker model generator is operative to generate the plurality of speaker models in the background by;

receiving training utterances from the plurality of speakers in the background, without prior knowledge of the speakers who spoke the respective training utterances;

blind clustering of the training utterances based on a predetermined criterion; and

training for each of the clusters a corresponding speaker model;

each of the models representing a speaker.