Detection of loudspeaker playback

1. A method of determining whether a sound has been generated by a loudspeaker, the method comprising:

receiving an audio signal representing at least a part of the sound;

separating the received audio signal into a first audio frequency component containing first audio frequencies and a second audio frequency component containing second audio frequencies;

obtaining a first statistical metric relating to sample values of the first audio frequency component and obtaining a second statistical metric relating to sample values of the second audio frequency component; and

based on said first statistical metric and said second statistical metric, determining whether said sound is indicative of use of a loudspeaker by detecting features resulting from a non-linearity that is greater at said first audio frequencies than at said second audio frequencies.

2. A method according to claim 1 , further comprising:

if said features indicative of use of a loudspeaker are identified, generating a loudspeaker indication output signal.

3. A method according to claim 1 , wherein at least one of the first statistical metric and the second statistical metric is a measure of an asymmetry of a waveform of the sound.

4. A method according to claim 1 , wherein at least one of the first statistical metric and the second statistical metric is a measure of skew of the sample values.

5. A method according to claim 1 , comprising:

obtaining a measure of skew of sample values of the first audio frequency component;

obtaining a measure of skew of sample values of the second audio frequency component; and

determining that the received audio signal has features resulting from a non-linearity that is greater at first audio frequencies than at second audio frequencies, if a difference between the measure of skew of sample values of the first audio frequency component and the measure of skew of sample values of the second audio frequency component exceeds a threshold value.

6. A method according to claim 1 , wherein obtaining the first statistical metric and obtaining the second statistical metric relating to sample values of the respective component comprises obtaining a first Cumulative Distribution Function relating to sample values of the first component and obtaining a second Cumulative Distribution Function relating to sample values of the second component.

7. A method according to claim 6 , wherein attempting to detect features resulting from a non-linearity that is greater at said first audio frequencies than at said second audio frequencies comprises:

comparing each Cumulative Distribution Function with a respective reference Cumulative Distribution Function for the respective component to obtain a measure of a difference therebetween; and

comparing said measures of difference obtained for the respective components.

8. A method according to claim 6 , wherein comparing signal content of different frequency bands comprises comparing the Cumulative Distribution Functions relating to the respective components.

9. A method according to claim 1 , comprising using a machine learning technique to attempt to detect in the received audio signal the features resulting from a non-linearity that is greater at low audio frequencies than at high audio frequencies.

10. A method according to claim 1 , comprising:

receiving a first signal representing the sound;

deriving from the first signal a second signal representing a part of the sound; and

attempting to detect in the second signal the features resulting from a non-linearity that is greater at first audio frequencies than at second audio frequencies.

11. A method according to claim 10 , wherein the second signal represents speech.

12. A method according to claim 11 , wherein the second signal represents voiced speech.

13. A method according to claim 10 , comprising:

receiving the first signal at a voice activity detector; and

deriving the second signal representing speech only if the voice activity detector detects speech in the first signal.

14. A method according to claim 1 , for use in a voice biometrics system, further comprising:

determining that the sound may result from a replay attack on the voice biometrics system, if it is determined that the sound was generated by a loudspeaker.

15. A system for determining whether a sound has been generated by a loudspeaker, the system comprising an input for receiving an audio signal representing at least a part of the sound; and a processor, and the system being configured for:

receiving the audio signal representing at least a part of the sound;

separating the received audio signal into a first audio frequency component containing first audio frequencies and a second audio frequency component containing second audio frequencies;

obtaining a first statistical metric relating to sample values of the first audio frequency component and obtaining a second statistical metric relating to sample values of the second audio frequency component; and

based on said first statistical metric and said second statistical metric, determining whether said sound is indicative of use of a loudspeaker by detecting features resulting from a non-linearity that is greater at said first audio frequencies than at said second audio frequencies.

16. A computer program product, comprising a non-transitory computer-readable medium having stored thereon instructions, the instructions, when read and executed by a processor, for performing a method of determining whether a sound has been generated by a loudspeaker, the method comprising:

receiving an audio signal representing at least a part of the sound;

separating the received audio signal into a first audio frequency component containing first audio frequencies and a second audio frequency component containing second audio frequencies;

obtaining a first statistical metric relating to sample values of the first audio frequency component and obtaining a second statistical metric relating to sample values of the second audio frequency component; and

based on said first statistical metric and said second statistical metric, determining whether said sound is indicative of use of a loudspeaker by detecting features resulting from a non-linearity that is greater at said first audio frequencies than at said second audio frequencies.