Voice authentication system and method

• US 9,491,167 B2
• Filed: 09/10/2013
• Issued: 11/08/2016
• Est. Priority Date: 09/11/2012
• Status: Active Grant

• Current Assignee: Auraya Pty Limited
• Original Assignee: Auraya Pty Limited
• Inventors: Talhami, Habib Emile, Malegaonkar, Amit Sadanand, Malegaonkar, Renuka Amit, Summerfield, Clive David
• Primary Examiner(s): Desrosiers, Evans
• Application Number:
  US14/427,571
• Publication Number:
  US 20150249664A1
• Time in Patent Office: 1,155 Days
• Field of Search: 713168-174, 713182-186, 713/202, 709/206, 709/225, 709/229, 709/249, 709/389, 726 2- 8
• US Class Current: 1/1
• CPC Class Codes: G06F 16/683 using metadata automatically derived from the content
• CPC Class Codes: G06F 16/9535 Search customisation based on user profiles and personalisation
• CPC Class Codes: G06F 21/32 using biometric data, e.g. fingerprints, iris scans or voiceprints
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: H04L 63/0861 using biometrical features, e.g. fingerprint, retina-scan cryptographic mechanisms or cryptographic arrangements for entity authentication using biological data H04L9/3231

Abstract

A method for maintaining a voiceprint utilizable to authenticate a user comprises associating a trainable universal background model with the user. A voiceprint for the user is generated utilizing the trainable universal background model and at least one enrollment voice sample provided by the user. The trainable universal background model is subsequently trained with additional voice samples provided by the user post enrollment that satisfy predefined adaptation criteria. The voiceprint is regenerated responsive to a change to the trainable universal background model.

Claims

1. A method for maintaining a voiceprint utilisable to authenticate a user, the method comprising:

selecting a universal background model that matches one or more acoustic parameters determined for the user and storing the selected universal background model in a database in association with a user profile for subsequent personalisation thereof;

generating a voiceprint for the user utilising the stored universal background model associated with the user and at least one enrolment voice sample provided by the user;

training the stored universal background model with additional voice samples provided by the user post enrolment that satisfy predefined adaptation criteria, wherein a first of the adaptation criteria comprises that the additional voice samples have been authenticated during a post enrolment authentication process referencing the user'"'"'s voiceprint such that, once associated with the user, the stored universal background model is only trained with additional authenticated voice samples provided by the user so as to exclusively personalise the stored universal background model to the user and wherein a second of the adaptation criteria comprises that the additional voice samples meet a predefined quality measure; and

re-generating the voiceprint responsive to a change to the stored universal background model.

2. A method in accordance with claim 1, wherein the initially selected universal background model corresponds to a seed universal background model which is built from voice samples provided by a cohort of speakers with one or more acoustic parameters shared by the user.

3. A method in accordance with claim 1, wherein the training step is carried out dynamically as new voice samples are provided for authentication.

4. A method in accordance with claim 1, wherein the quality measure of the second adaptation criteria is met when an individual equal error rate for the additional voice samples is less than a predefined threshold.

5. A method in accordance with claim 1, further comprising continuing to train the stored universal background model with additional voice samples until a quality measure for at least one of the stored universal background model and voiceprint has been met.

6. A voice authentication system comprising:

a processor coupled to a memory operable to generate a voiceprint for a user, utilising a stored universal background model associated with the user and at least one enrolment voice sample provided by the user, the universal background model being initially selected from a set of different universal background models based on one or more acoustic parameters determined for the user and wherein the selected universal background model is subsequently stored in a database in association with a user profile for subsequent personalisation thereof;

a biometric trainer operable to train the stored universal background model with an additional voice sample provided by the user post enrolment that satisfies predefined adaptation criteria, wherein a first of the adaptation criteria comprises that the additional voice samples have been authenticated during a post enrolment authentication process referencing the user'"'"'s voiceprint such that, once associated with the user, the stored universal background model is only trained with additional authenticated voice samples provided by the user so as to exclusively personalise the stored universal background model to the user and wherein a second of the adaptation criteria comprises that the additional voice samples meet a predefined quality measure and wherein the processing module is further operable to regenerate the voiceprint responsive to a change to the stored universal background model.

7. A voice authentication system in accordance with claim 6, wherein the voiceprint is regenerated utilising the additional voice sample.

8. A voice authentication system in accordance with claim 6, wherein the initially selected universal background model corresponds to a seed universal background model which is built from voice samples provided by a cohort of speakers with one or more shared voice characteristics.

9. A voice authentication system in accordance with claim 6, wherein the quality measure of the second adaptation criteria is met when an individual equal error rate for the additional voice samples is less than a predefined threshold.

10. A voice authentication system in accordance with claim 6, wherein the biometric trainer is operable to continue to train the stored universal background model with additional voice samples until a quality measure for at least one of the stored universal background model and voiceprint has been met.

11. Computer program code residing on a non-transitory medium comprising at least one instruction which, when executed by a computer, is arranged to implement a method for maintaining a voiceprint, comprising:

selecting a universal background model that matches one or more acoustic parameters determined for the user and storing the selected universal background model in a database in association with a user profile for subsequent personalisation thereof;

generating a voiceprint for the user utilising the stored universal background model associated with the user and at least one enrolment voice sample provided by the user;

training the stored universal background model with additional voice samples provided by the user post enrolment that satisfy predefined adaptation criteria, a first of the adaptation criteria comprises that the voice samples have been authenticated during a post enrolment authentication process referencing the user'"'"'s voiceprint such that, once associated with the user, the stored universal background model is only trained with additional authenticated voice samples provided by the user so as to exclusively personalise the stored universal background model to the user and a second of the adaptation criteria comprising that the additional voice samples meet a predefined quality measure; and

re-generating the voiceprint responsive to a change to the stored universal background model.