Method and apparatus for detecting splicing attacks on a speaker verification system

1. A method of detecting an occurrence of splicing in a test speech signal, comprising:

comparing one or more signal discontinuities in the test speech signal to one or more reference speech signals corresponding to the test speech signal; and

distinguishing, based on the comparing, between a naturally-occurring signal discontinuity and a splicing-based signal discontinuity.

2. The method of claim 1 , further comprising:

calculating a frame-based spectral-like representation S T of the speech signal;

calculating a frame-based spectral-like representation S E of a reference speech signal corresponding to the speech signal;

aligning S T and S E in time and frequency;

calculating a distance function associated with aligned S T and S E ;

evaluating the distance function to determine a score; and

comparing the score to a threshold to detect if splicing occurs in the speech signal.

3. The method of claim 2 , wherein the frame-based spectral-like representations are short time Fourier transforms.

4. The method of claim 2 , wherein the frame-based spectra-like representations are Mel-frequency cepstral coefficients.

5. The method of claim 2 , wherein performing a time alignment between S T and S E includes performing a dynamic time warping.

6. The method of claim 2 , wherein performing a spectral alignment between S T and S E includes calculating a global linear transformation of S E so as to minimize a difference between S T and the global linear transformation of S E .

7. The method of claim 2 , wherein calculating the distance function further comprises:

(i) constructing one or more models that simulates a difference S T −S E ;

(ii) selecting a prime model being one of the one or more models that predicts an actual difference S T −S E better than others of the one or more models;

(iii) using the prime model to extrapolate a predicted difference S Tp −S Ep ;

(iv) calculating a difference between the difference S T −S E and the predicted difference S Tp −S Ep .

8. The method of claim 7 , wherein the one or more models includes at least one of a 0 th order model, a first order model and a second order model.

9. The method of claim 7 , wherein evaluating the distance function to determine a score includes at least one of (i) assigning an n th highest value of the distance function as the score, (ii) assigning an average of m values of the distance function as the score and (iii) assigning a median of p values of the distance function as the score, where n, m and p are integers.

10. The method of claim 1 , wherein the reference speech signal represents to a predetermined phrase spoken by a presumed speaker of the speech signal.

11. The method of claim 1 , wherein the test speech signal and the corresponding reference speech signal are speech segments.

12. An apparatus for detecting an occurrence of splicing in a speech signal, comprising:

a processor; and

a memory configured to store instructions to be executed by the processor;

the processor being configured to execute the instructions thereby causing the apparatus to: (i) compare one or more signal discontinuities in the test speech signal to one or more reference speech signals corresponding to the test speech signal, and (ii) distinguish, based on the compare, between a naturally-occurring signal discontinuity and a splicing-based signal discontinuity.

13. The apparatus of claim 12 , the processor being further configured to execute the instructions thereby causing the apparatus to:

calculate a frame-based spectral-like representation S T of the speech signal;

calculate a frame-based spectral-like representation S E of a reference speech signal corresponding to the speech signal;

align S T and S E in time and frequency;

calculate a distance function associated with aligned S T and S E ;

evaluate the distance function to determine a score; and

compare the score to a threshold to detect if splicing occurs in the speech signal.

14. The apparatus of claim 13 , wherein the frame-based spectral-like representations are short time Fourier transforms.

15. The apparatus of claim 13 , wherein the frame-based spectra-like representations are Mel-frequency cepstral coefficients.

16. The apparatus of claim 13 , wherein performing a time alignment between S T and S E includes performing a dynamic time warping.

17. The apparatus of claim 13 , wherein performing a spectral alignment between S T and S E includes calculating a global linear transformation of S E so as to minimize a difference between S T and the global linear transformation of S E .

18. The apparatus of claim 13 , wherein calculating the distance function further comprises:

(i) constructing one or more models that simulates a difference S T −S E ;

(ii) selecting a prime model being one of the one or more models that predicts an actual difference S T −S E better than others of the one or more models;

(iii) using the prime model to extrapolate a predicted difference S Tp −S Ep ;

(iv) calculating a difference between the difference S T −S E and the predicted difference S Tp −S Ep .

19. The apparatus of claim 18 , wherein the one or more models includes at least one of a 0th order model, a first order model and a second order model.

20. The apparatus of claim 12 , wherein the reference speech signal represents to a predetermined phrase spoken by a presumed speaker of the speech signal.

21. The apparatus of claim 12 , wherein the test speech signal and the corresponding reference speech signal are speech segments.

22. A non-transitory computer-readable medium with computer code instruction stored thereon, the computer code instructions when executed by an a processor cause an apparatus to: (i) compare one or more signal discontinuities in the test speech signal to one or more reference speech signals corresponding to the test speech signal, and (ii) distinguish, based on the compare, between a naturally-occurring signal discontinuity and a splicing-based signal discontinuity.

23. The non-transitory computer-readable medium of claim 22 , the computer code instructions when executed by an a processor further cause an apparatus to:

calculate a frame-based spectral-like representation ST of the speech signal;

calculate a frame-based spectral-like representation SE of a reference speech signal corresponding to the speech signal;

align ST and SE in time and frequency;

calculate a distance function associated with aligned ST and SE;

evaluate the distance function to determine a score; and

compare the score to a threshold to detect if splicing occurs in the speech signal.