Speaker verification and speaker identification based on eigenvoices

• US 6,141,644 A
• Filed: 09/04/1998
• Issued: 10/31/2000
• Est. Priority Date: 09/04/1998
• Status: Expired due to Term

• Current Assignee: Panasonic Intellectual Property Corporation of America (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Junqua, Jean-Claude, Boman, Robert, Nguyen, Patrick, Kuhn, Roland
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): CHAWAN, VIJAY B
• Application Number:
  US09/148,911
• Time in Patent Office: 788 Days
• Field of Search: 704/238, 704/240, 704/243, 704/272, 704/246, 704/250, 704/270, 704/273, 704/236
• US Class Current: 704/273
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction

Abstract

Speech models are constructed and trained upon the speech of known client speakers (and also impostor speakers, in the case of speaker verification). Parameters from these models are concatenated to define supervectors and a linear transformation upon these supervectors results in a dimensionality reduction yielding a low-dimensional space called eigenspace. The training speakers are then represented as points or distributions in eigenspace. Thereafter, new speech data from the test speaker is placed into eigenspace through a similar linear transformation and the proximity in eigenspace of the test speaker to the training speakers serves to authenticate or identify the test speaker.

Claims

1. A method for assessing speech with respect to a predetermined client speaker, comprising:

training a set of speech models upon the speech from a plurality of training speakers, the plurality of training speakers including at least one client speaker;

constructing an eigenspace to represent said plurality of training speakers by performing dimensionality reduction upon said set of models to generate a set of basis vectors that define said eigenspace;

representing said client speaker as a first location in said eigenspace;

processing new speaker input data by training a new speech model upon said input data and by performing dimensionality reduction upon said new speech model to generate a representation of said new speaker as a second location in eigenspace;

assessing the proximity between said first and second locations and using said assessment as an indication of whether the new speaker is the client speaker.

2. A speaker identification method according claim 1 wherein said plurality of training speakers includes a plurality of different client speakers and wherein said method further comprises:

representing each of said plurality of client speakers as training speaker locations in said eigenspace, and

assessing the proximity between said second location and said training speaker locations and identifying said new speaker as a selected one of said plurality of client speakers based at least in part on said proximity assessment.

3. A speaker verification method according to claim 1 wherein said plurality of training speakers includes at least one impostor speaker that is represented as a third location in eigenspace.

4. A speaker verification method according to claim 3 further comprising additionally assessing the proximity between said second and third locations and using said additional assessment as a further indication of whether the new speaker is the client speaker.

5. The method of claim 1 wherein said step of assessing proximity is performed by determining the distance between said first and second locations.

6. The method of claim 1 wherein said training speakers are represented as locations in said eigenspace.

7. The method of claim 1 wherein said training speakers are represented as points in said eigenspace.

8. The method of claim 1 wherein said training speakers are represented as distributions in said eigenspace.

9. The method of claim 1 wherein said step of processing new speaker input data includes using said input data to generate a probability function and then maximizing said probability function to determine a maximum likelihood vector that lies within said eigenspace.

10. The method of claim 1 wherein said plurality of training speakers includes a plurality of client speakers and at least one impostor speaker.

11. The method of claim 1 further comprising periodically assessing the proximity between said first and second locations and using said assessment as an indication of whether the new speaker is the client speaker to determine if the identity of said new speaker changes.