Tamper-resistant element for use in speaker recognition

Patent metadata

Field	Value
Publication number	US-9431016-B2
Application number	US-201414281373-A
Country	US
Kind code	B2
Filing date	May 19, 2014
Priority date	Jan 17, 2014
Publication date	Aug 30, 2016
Grant date	Aug 30, 2016

Abstract

A tamper-resistant element for use in speaker recognition, the tamper-resistant element being adapted for storing data representing speaker information based on speaker recognition enrollment data and for checking whether information based on a speaker recognition testing signal matches the speaker information. The tamper-resistant element is also adapted for carrying out a data integrity check. Also, a system including such a tamper-resistant element and method for speaker recognition.

First claim

The invention claimed is:

1. A speaker recognition system, comprising:
a tamper-resistant element, comprising a first processor and a memory; and
a second processor, provided separate from the tamper-resistant element;
wherein:
the second processor is connected to a source of audio samples, and the second processor is provided with a Universal Background Model for speaker recognition;
the second processor is configured such that, during an enrollment phase, samples received from an enrolling user are compared with the Universal Background Model in the second processor to form data representing speaker information for the enrolling user;
the tamper-resistant element is configured for receiving the data representing speaker information for an enrolling user from the second processor, and the memory of the tamper-resistant element is configured for storing said data representing speaker information for the enrolling user, such that the enrolling user becomes an enrolled user; and
the second processor is configured such that, during a verification phase, samples received from a speaker subject to verification are compared with the Universal Background Model in the second processor to form data representing speaker information for the speaker subject to verification; and

wherein the tamper-resistant element is configured for receiving the data representing speaker information for the speaker subject to verification from the second processor, and the processor of the tamper-resistant element is configured for comparing the data representing speaker information for the speaker subject to verification with the data representing speaker information for the enrolled user, enabling a determination based on said comparison whether the speaker subject to verification is the enrolled user.

2. A speaker recognition system as claimed in claim 1 , wherein the processor of the tamper-resistant element is configured for making said determination based on said comparison whether the speaker subject to verification is the enrolled user.

3. A speaker recognition system as claimed in claim 2 , wherein the processor of the tamper-resistant element is configured for determining based on said comparison whether the speaker subject to verification is the enrolled user by calculating a likelihood that the speaker subject to verification is the enrolled user.

4. A speaker recognition system as claimed in claim 1 , wherein the tamper-resistant element is configured for outputting a result of the comparison of the data representing speaker information for the speaker subject to verification with the data representing speaker information for the enrolled user.

5. A speaker recognition system as claimed in claim 1 , wherein the tamper-resistant element is configured for performing a data integrity check on the received data representing speaker information for the enrolling user, and/or on the received data representing speaker information for the speaker subject to verification.

6. A speaker recognition system as claimed in claim 1 , wherein:
the memory of the tamper-resistant element is configured for storing an anti-spoofing model;
the second processor is configured such that, during a verification phase, samples received from the speaker subject to verification are processed to obtain data for use in a spoof check, wherein the data for use in a spoof check is distinct from the data representing speaker information for the speaker subject to verification;
the tamper-resistant element is configured for receiving the data for use in a spoof check; and
the processor of the tamper-resistant element is configured for comparing the data for use in a spoof check with the anti-spoofing model, enabling a determination based on said comparison whether the received samples represent a spoof attack.

7. A speaker recognition system as claimed in claim 6 , wherein the tamper-resistant element is configured for performing a data integrity check on the received data for use in the spoof check.

8. A speaker recognition system as claimed in claim 6 , wherein the processor of the tamper-resistant element is configured for making said determination based on said comparison whether the received samples represent a spoof attack.

9. A speaker recognition system as claimed in claim 6 , wherein the tamper-resistant element is configured for outputting a result of the comparison of the data for use in a spoof check with the anti-spoofing model.

10. A speaker recognition system as claimed in claim 1 , wherein the data representing speaker information for the enrolling user and the data representing speaker information for the speaker subject to verification comprise Probabilistic Linear Discriminant Analysis statistics.

11. A speaker recognition system as claimed in claim 1 , wherein the data representing speaker information for the enrolling user and the data representing speaker information for the speaker subject to verification comprise feature vectors derived from received audio samples.

12. A speaker recognition system as claimed in claim 11 , wherein the feature vectors comprise Mel-frequency cepstral coefficients.

13. A speaker recognition system as claimed in claim 11 , wherein the Universal Background Model is a Hidden Markov Model, and wherein the feature vectors comprise Mel-frequency cepstral coefficient statistics for a subset of Gaussians of the Hidden Markov Model Universal Background Model.

14. A speaker recognition system as claimed in claim 1 , wherein the Universal Background Model is a Hidden Markov Model, and wherein it is assumed that the optimal sequences of states after Viterbi decoding for the Hidden Markov Model Universal Background Model and for the Hidden Markov Model representing speaker information are identical.

15. A speaker recognition system as claimed in claim 5 , wherein the data integrity check comprises an injection integrity check.

16. A speaker recognition system as claimed in claim 15 , wherein the tamper-resistant element is configured for performing the injection integrity check on the received data representing speaker information for the enrolling user, and/or on the received data representing speaker information for the speaker subject to verification, by cryptographic means.

17. A speaker recognition system as claimed in claim 5 , wherein the data integrity check comprises a speaker recognition knowledge based integrity check.

18. A speaker recognition system as claimed in claim 17 ,
wherein the data representing speaker information for the enrolling user and the data representing speaker information for the speaker subject to verification comprise Probabilistic Linear Discriminant Analysis statistics, and
wherein the tamper-resistant element is configured for performing the speaker recognition knowledge based integrity check on the received data representing speaker information for the enrolling user, and/or on the received data representing speaker information for the speaker subject to verification, by testing properties of the Probabilistic Linear Discriminant Analysis statistics for the received data representing speaker information for the speaker subject to verification.

19. A speaker recognition system as claimed in claim 17 ,
wherein the Universal Background Model is a Hidden Markov Model,
wherein the data representing speaker information for the enrolling user and the data representing speaker information for the speaker subject to verification comprise Mel-frequency cepstral coefficient statistics for a subset of Gaussians of t

Assignees

• Cirrus Logic Int Semiconductor Ltd

Inventors

• Aviles-Casco Carlos Vaquero
• Rodriguez Luiz Buera
• Gomar Marta Garcia

Classifications

• G10L15/14

  using statistical models, e.g. Hidden Markov Models [HMMs] (G10L15/18 takes precedence) · CPC title

• G10L17/00

  Speaker identification or verification techniques · CPC title

• G10L15/142

  Hidden Markov Models [HMMs] · CPC title

• G10L17/02 Primary

  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 · CPC title

• G06F21/32

  using biometric data, e.g. fingerprints, iris scans or voiceprints · CPC title