Anti-spoofing

• US 9,767,806 B2
• Filed: 09/24/2014
• Issued: 09/19/2017
• Est. Priority Date: 09/24/2013
• Status: Active Grant

• Current Assignee: Cirrus Logic Incorporated
• Original Assignee: Cirrus Logic International Semiconductor Ltd. (Cirrus Logic Incorporated)
• Inventors: Gimnez, Alfonso Ortega, Rodriguez, Luis Buera, Avils-Casco, Carlos Vaquero
• Primary Examiner(s): Sharma, Neeraj
• Application Number:
  US14/495,391
• Publication Number:
  US 20150088509A1
• Time in Patent Office: 1,091 Days
• Field of Search: 704243, 704205, 704234, 704246, 704256, 704247, 704270, 704249, 341200, 381 941, 700 94, 726 19, 382115
• 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
• CPC Class Codes: G10L 17/04 Training, enrolment or model building
• CPC Class Codes: G10L 17/22 Interactive procedures; Man-machine interfaces

Abstract

System for classifying whether audio data received in a speaker recognition system is genuine or a spoof using a Gaussian classifier and method for classifying whether audio data received in a speaker recognition system is genuine or a spoof using a Gaussian classifier.

Claims

1. A speaker recognition system adapted for receiving audio data, the system being adapted for:

receiving audio data under test;

obtaining a Medium Frequency Relative Energy (MF) parameter, comprising a ratio between an energy of the received audio data under test in a predetermined frequency band and an energy of a complete frequency spectrum of the received audio data under test; and

classifying using a Gaussian classifier whether the received audio data under test is genuine or represents a recording replayed through a loudspeaker, based on the Medium Frequency Relative Energy (MF) parameter, wherein the Gaussian classifier is trained by the following steps;

a first Gaussian is obtained by;

receiving genuine audio data;

obtaining a first Medium Frequency Relative Energy (MF) parameter, comprising the ratio between the energy of the genuine audio data in a predetermined frequency band and the energy of the complete frequency spectrum of the genuine audio data;

receiving audio data representing recordings replayed through a loudspeaker; and

modelling the genuine audio data;

and wherein a second Gaussian is obtained by;

receiving audio data representing recordings replayed through a loudspeaker;

obtaining a second Medium Frequency Relative Energy (MF) parameter, comprising the ratio between the energy of the audio data representing recordings replayed through a loudspeaker in a predetermined frequency band and the energy of the complete frequency spectrum of the audio data representing recordings replayed through a loudspeaker; and

modelling the audio data representing recordings replayed through a loudspeaker with a second Gaussian.

2. A speaker recognition system according to claim 1, wherein a lower end of the predetermined frequency band is in a range from 30 Hz to 150 Hz.

3. A speaker recognition system according to claim 1, wherein an upper end of the predetermined frequency band is in a range from 150 Hz to 270 Hz.

4. A speaker recognition system according to claim 1, wherein the considered parameters of the audio data further comprise a feature vector distance.

5. A speaker recognition system according to claim 4, wherein the feature vector distance is calculated with regard to a constant value.

6. A speaker recognition system according to claim 4, wherein the feature vector distance is calculated with regard to average feature vectors derived from enrolment data.

7. A speaker recognition system according to claim 1, wherein the considered parameters of the audio data further comprise a spectral ratio.

8. A speaker recognition system according to claim 1, wherein new parameters for the Gaussian classifier are found by adaptation of previous parameters of the Gaussian classifier using adaptation audio data.

9. A speaker recognition system according to claim 8, wherein the number of available samples of adaptation audio data is considered in the adaptation process.

10. A speaker recognition system according to claim 8, wherein mean vector(s) and/or covariance matrices and/or an a priori probability of one, two, three, four or more Gaussians representing the region of audio data parameters from genuine audio data and/or wherein mean vector(s) and/or the covariance matrices and/or an a priori probability of one, two, three, four or more Gaussians representing the region of audio data parameters from audio data representing recordings replayed through loudspeakers are adapted.

11. A speaker recognition system according to claim 8, wherein enrollment audio data comprises the adaptation audio data.

12. A speaker recognition system according to claim 8, wherein the adaptation audio data comprises genuine audio data and/or audio data representing a recording replayed through a loudspeaker.

13. A speaker recognition system according to claim 8, wherein the adaptation audio data is chosen depending on information that the Gaussian classifier should model.

14. A method in a speaker recognition system for classifying whether audio data is genuine or represents a recording replayed through a loudspeaker, the method comprising:

receiving the audio data, and

classifying using a Gaussian classifier whether the received audio data is genuine or represents a recording replayed through a loudspeaker, wherein Gaussians are used to model a region of audio data parameters from genuine audio data and wherein Gaussians are used to model a region of audio data parameters from audio data representing recordings replayed through loudspeakers, based on a Medium Frequency Relative Energy (MF) parameter, and

wherein;

the Medium Frequency Relative Energy (MF) parameter comprises a ratio between an energy of the audio data in a predetermined frequency band and an energy of a complete frequency spectrum of the audio data; and

the Gaussian classifier is trained by the following steps;

a first Gaussian is obtained by;

receiving genuine audio data;

obtaining a first Medium Frequency Relative Energy (MF) parameter, comprising the ratio between the energy of the genuine audio data in a predetermined frequency band and the energy of the complete frequency spectrum of the genuine audio data;

receiving audio data representing recordings replayed through a loudspeaker; and

modelling the genuine audio data;

and wherein a second Gaussian is obtained by;

receiving audio data representing recordings replayed through a loudspeaker;

obtaining a second Medium Frequency Relative Energy (MF) parameter, comprising the ratio between the energy of the audio data representing recordings replayed through a loudspeaker in a predetermined frequency band and the energy of the complete frequency spectrum of the audio data representing recordings replayed through a loudspeaker; and

modelling the audio data representing recordings replayed through a loudspeaker with a second Gaussian.

15. A non-transitory computer-readable medium comprising computer-readable instructions that, when executed on a computer, are adapted to carry out a method according to claim 14.

16. A speaker recognition system adapted for receiving audio data, the system being adapted for:

receiving audio data under test;

obtaining a Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, comprising a ratio between an energy of the received audio data under test in a predetermined frequency band and an energy of a complete frequency spectrum of the received audio data under test; and

classifying using a Gaussian classifier whether the received audio data under test is genuine or represents a recording replayed through a loudspeaker, based on the Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, wherein the Gaussian classifier is trained by the following steps;

a first Gaussian is obtained by;

receiving genuine audio data;

obtaining a first Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, comprising the ratio between the energy of the genuine audio data in a predetermined frequency band and the energy of the complete frequency spectrum of the genuine audio data;

receiving audio data representing recordings replayed through a loudspeaker; and

modelling the genuine audio data;

and wherein a second Gaussian is obtained by;

receiving audio data representing recordings replayed through a loudspeaker;

obtaining a Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, comprising the ratio between the energy of the audio data representing recordings replayed through a loudspeaker in a predetermined frequency band and the energy of the complete frequency spectrum of the audio data representing recordings replayed through a loudspeaker; and

modelling the audio data representing recordings replayed through a loudspeaker with a second Gaussian.

17. A speaker recognition system according to claim 16, wherein the predetermined cut-off frequency is lower than 750 Hz.

18. A speaker recognition system according to claim 16, wherein the considered parameters of the audio data further comprise a feature vector distance.

19. A speaker recognition system according to claim 18, wherein the feature vector distance is calculated with regard to a constant value.

20. A speaker recognition system according to claim 18, wherein the feature vector distance is calculated with regard to average feature vectors derived from enrolment data.

21. A speaker recognition system according to claim 18, wherein the considered parameters of the audio data further comprise a spectral ratio.

22. A speaker recognition system according to claim 16, wherein new parameters for the Gaussian classifier are found by adaptation of previous parameters of the Gaussian classifier using adaptation audio data.

23. A speaker recognition system according to claim 22, wherein the number of available samples of adaptation audio data is considered in the adaptation process.

24. A speaker recognition system according to claim 22, wherein mean vector(s) and/or covariance matrices and/or an a priori probability of one, two, three, four or more Gaussians representing the region of audio data parameters from genuine audio data and/or wherein mean vector(s) and/or the covariance matrices and/or an a priori probability of one, two, three, four or more Gaussians representing the region of audio data parameters from audio data representing recordings replayed through loudspeakers are adapted.

25. A speaker recognition system according to claim 22, wherein enrollment audio data comprises the adaptation audio data.

26. A speaker recognition system according to claim 22, wherein the adaptation audio data comprises genuine audio data and/or audio data representing a recording replayed through a loudspeaker.

27. A speaker recognition system according to claim 22, wherein the adaptation audio data is chosen depending on the information that the Gaussian classifier should model.

30. A speaker recognition system according to claim 16, wherein the predetermined cut-off frequency is between 250 Hz and 750 Hz.

28. A method in a speaker recognition system for classifying whether audio data is genuine or represents a recording replayed through a loudspeaker, the method comprising:

receiving the audio data, and

classifying using a Gaussian classifier whether the received audio data is genuine or represents a recording replayed through a loudspeaker, wherein Gaussians are used to model a region of audio data parameters from genuine audio data and wherein Gaussians are used to model a region of audio data parameters from audio data representing recordings replayed through loudspeakers, based on the Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, and

wherein;

the Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter comprises 1, 2, 3 or more or all LF-MFCC extracted from a region of the audio data having frequencies lower than a predetermined cut-off frequency; and

wherein the Gaussian classifier is trained by the following steps;

a first Gaussian is obtained by;

receiving genuine audio data;

obtaining a first Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, comprising the ratio between the energy of the genuine audio data in a predetermined frequency band and the energy of the complete frequency spectrum of the genuine audio data;

receiving audio data representing recordings replayed through a loudspeaker; and

modelling the genuine audio data;

and wherein a second Gaussian is obtained by;

receiving audio data representing recordings replayed through a loudspeaker;

obtaining a Low Frequency Mel Frequency Cepstral Coefficients (LF-MFCC) parameter, comprising the ratio between the energy of the audio data representing recordings replayed through a loudspeaker in a predetermined frequency band and the energy of the complete frequency spectrum of the audio data representing recordings replayed through a loudspeaker; and

modelling the audio data representing recordings replayed through a loudspeaker with a second Gaussian.

29. A non-transitory computer-readable medium comprising computer-readable instructions that, when executed on a computer, are adapted to carry out a method according to claim 28.