Detecting deep-fake audio through vocal tract reconstruction

1. An apparatus comprising at least one processor and at least one memory including computer program code, the at least one memory and computer program code configured to, with the processor, cause the apparatus to at least:

generate a model of a vocal tract based at least in part on frequency response of one or more organic audio samples from a user;

identify a set of bigram-feature pairs from the one or more organic audio samples;

estimate, from the generated model of the vocal tract, the cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs;

receive a candidate audio sample;

identify bigram-feature pairs of the candidate audio sample that are in the set of bigram-feature pairs;

calculate a cross-sectional area of a theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample; and

identify the candidate audio sample as a deep-fake audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample failing to correspond within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

2. The apparatus of claim 1 , wherein the apparatus is further caused to:

identify the candidate audio sample as an organic audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample corresponding within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

3. The apparatus of claim 1 , wherein causing the apparatus to identify a set of bigram-feature pairs from the one or more audio samples comprises causing the apparatus to identify bigram-feature pairs based, at least in part, on an estimation of the cross-sectional area of a vocal tract of a user when speaking respective bigram-feature pairs.

4. The apparatus of claim 1 , wherein causing the apparatus to identify a set of bigram-feature pairs from the one or more audio samples comprises causing the apparatus to identify a set of bigram-feature pairs based, at least in part, on divergences in distributions of features in specific bigrams between deep-fake audio and organic audio.

5. The apparatus of claim 4 , wherein divergences in distributions of features in specific bigrams between deep-fake audio and organic audio are established based at least in part on a degree of overlap between probability density function graphs of the specific bigrams for deep-fake audio samples and organic audio samples.

6. The apparatus of claim 1 , wherein causing the apparatus to generate the model of the vocal tract based at least in part on frequency response of one or more organic audio samples from a user comprises causing the apparatus to:

process the one or more organic audio samples through a Fast Fourier Transform to obtain relative amplitudes for frequencies of a voice of the user.

7. The apparatus of claim 1 , wherein causing the apparatus to identify the candidate audio sample as a deep-fake audio sample further comprises causing the apparatus to:

provide an explanation of how the candidate audio sample was identified as a deep-fake audio sample.

8. A method comprising:

generating a model of a vocal tract based at least in part on frequency response of one or more organic audio samples from a user;

identifying a set of bigram-feature pairs from the one or more organic audio samples;

estimating, from the generated model of the vocal tract, the cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs;

receiving a candidate audio sample;

identifying bigram-feature pairs of the candidate audio sample that are in the set of bigram-feature pairs;

calculating a cross-sectional area of a theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample; and

identifying the candidate audio sample as a deep-fake audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample failing to correspond within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

9. The method of claim 8 , further comprising:

identifying the candidate audio sample as an organic audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample corresponding within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

10. The method of claim 8 , wherein identifying a set of bigram-feature pairs from the one or more audio samples comprises identifying bigram-feature pairs based, at least in part, on an estimation of the cross-sectional area of a vocal tract of a user when speaking respective bigram-feature pairs.

11. The method of claim 8 , wherein identifying a set of bigram-feature pairs from the one or more audio samples comprises identifying a set of bigram-feature pairs based, at least in part, on divergences in distributions of features in specific bigrams between deep-fake audio and organic audio.

12. The method of claim 11 , wherein divergences in distributions of features in specific bigrams between deep-fake audio and organic audio are established based at least in part on a degree of overlap between probability density function graphs of the specific bigrams for deep-fake audio samples and organic audio samples.

13. The method of claim 8 , wherein generating the model of the vocal tract based at least in part on frequency response of one or more organic audio samples from a user comprises:

processing the one or more organic audio samples through a Fast Fourier Transform to obtain relative amplitudes for frequencies of a voice of the user.

14. The method of claim 8 , wherein identifying the candidate audio sample as a deep-fake audio sample further comprises:

providing an explanation of how the candidate audio sample was identified as a deep-fake audio sample.

15. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-executable program code portions stored therein, the computer-executable program code portions comprising program code instructions configured to:

generate a model of a vocal tract based at least in part on frequency response of one or more organic audio samples from a user;

identify a set of bigram-feature pairs from the one or more organic audio samples;

estimate, from the generated model of the vocal tract, the cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs;

receive a candidate audio sample;

identify bigram-feature pairs of the candidate audio sample that are in the set of bigram-feature pairs;

calculate a cross-sectional area of a theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample; and

identify the candidate audio sample as a deep-fake audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample failing to correspond within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

16. The computer program product of claim 15 , further comprising program code instructions to:

identify the candidate audio sample as an organic audio sample in response to the calculated cross-sectional area of the theoretical vocal tract of a user when speaking the identified bigram-feature pairs of the candidate audio sample corresponding within a predefined measure of the estimated cross-sectional area of the vocal tract of the user when speaking the set of bigram-feature pairs.

17. The computer program product of claim 15 , wherein the program code instructions to identify a set of bigram-feature pairs from the one or more audio samples comprise program code instructions to identify bigram-feature pairs based, at least in part, on an estimation of the cross-sectional area of a vocal tract of a user when speaking respective bigram-feature pairs.

18. The computer program product of claim 15 , wherein the program code instructions to identify a set of bigram-feature pairs from the one or more audio samples comprise program code instructions to identify a set of bigram-feature pairs based, at least in part, on divergences in distributions of features in specific bigrams between deep-fake audio and organic audio.

19. The computer program product of claim 18 , wherein divergences in distributions of features in specific bigrams between deep-fake audio and organic audio are established based at least in part on a degree of overlap between probability density function graphs of the specific bigrams for deep-fake audio samples and organic audio samples.

20. The computer program product of claim 15 , wherein the program code instructions to identify the candidate audio sample as a deep-fake audio sample further comprise program code instructions to:

provide an explanation of how the candidate audio sample was identified as a deep-fake audio sample.