Deepfake detection
Disclosed are systems and methods including software processes executed by a server that detect audio-based synthetic speech (“deepfakes”) in a call conversation. The server applies an NLP engine to transcribe call audio and analyze the text for anomalous patterns to detect synthetic speech. Additionally or alternatively, the server executes a voice “liveness” detection system for detecting machine speech, such as synthetic speech or replayed speech. The system performs phrase repetition detection, background change detection, and passive voice liveness detection in call audio signals to detect liveness of a speech utterance. An automated model update module allows the liveness detection model to adapt to new types of presentation attacks, based on the human provided feedback.
1 . A computer-implemented method for detecting machine-based speech in calls, the method comprising:
in response to detecting a first speech region in a first audio signal received during a call from a caller, applying, by a computer, an event classifier on a first set of acoustic features extracted from the first audio signal to classify a first set of background audio events in the first audio signal;
in response to detecting a second speech region in a second audio signal received during the call, applying, by the computer, the event classifier on a second set of acoustic features extracted from the second audio signal to classify a second set of background audio events in the second audio signal;
generating, by the computer, a machine likelihood score based upon an amount of similarity between the first set of background audio events and the second set of background audio events obtained during the call; and
identifying, by the computer, the caller as a machine caller in response to determining that the machine likelihood score satisfies a similarity threshold.
2 . The method of claim 1 , wherein generating the machine likelihood score further comprises generating the machine likelihood score based upon identifying a first background event in the first audio signal and a second background event in the second audio signal within a time period during the call, whereby the machine caller provides a plurality of sets of background audio events that are distinct throughout the call.
3 . The method of claim 1 , wherein generating the machine likelihood score further comprises applying an attack prediction model to the first set of background events and the second set of background events to generate the machine likelihood score over a first segment corresponding to the first audio signal and a second segment corresponding to the second audio signal.
4 . The method of claim 1 , wherein applying the event classifier on the first set of acoustic features includes applying, by the computer, the event classifier to detect one or more first anomalous events within the first audio signal; and
wherein applying the event classifier on the second set of acoustic features includes applying, by the computer, the event classifier to detect one or more second anomalous events within the second audio signal.
5 . The method of claim 1 , further comprising:
applying, by the computer, an estimation model to the first set of acoustic features to determine a first quality metric indicating a first degree of change in quality through the first audio signal; and
applying, by the computer, the estimation model to the second set of acoustic features to determine a second quality metric indicating a second degree of change in quality through the second audio signal.
6 . The method of claim 5 , wherein the first quality metric and the second quality metric each include at least one of: (i) a respective noise type, (ii) a respective reverberation ratio, or (iii) a respective signal to noise (SNR) ratio.
7 . The method of claim 5 , wherein generating the machine likelihood score includes generating, by the computer, the machine likelihood score based upon a comparison between the first quality metric and the second quality metric.
8 . The method of claim 1 , further comprising:
extracting, by the computer, the first set of acoustic features from a first background in the first audio signal; and
extracting, by the computer, the second set of acoustic features from a second background in the second audio signal.
9 . The method of claim 1 , further comprising identifying, by the computer, the caller as a human caller in response to determining that the machine likelihood score does not satisfy the similarity threshold.
10 . The method of claim 1 , further comprising generating, by the computer, an indication, for a user interface, of the caller as a machine caller in response to determining that the machine likelihood score satisfies the threshold.
11 . A system for detecting machine-based speech in calls, comprising:
a computer comprising one or more processors and configured to:
detect a first speech region in a first audio signal received during a call from a caller;
in response to the computer detecting the first speech region in the first audio signal received during the call from the caller, apply an event classifier on a first set of acoustic features extracted from the first audio signal to classify a first set of background audio events in the first audio signal;
in response to detecting a second speech region in a second audio signal received during the call, apply the event classifier on a second set of acoustic features extracted from the second audio signal to classify a second set of background audio events in the second audio signal;
generate a machine likelihood score based upon an amount of similarity between the first set of background audio events and the second set of background audio events obtained during the call; and
identify the caller as a machine caller in response to determining that the machine likelihood score satisfies a similarity threshold.
12 . The system of claim 11 , wherein, when generating the machine likelihood score, the computer is further configured to generate the machine likelihood score based upon identifying a first background event in the first audio signal and a second background event in the second audio signal within a time period during the call, whereby the machine caller provides a plurality of sets of background audio events that are distinct throughout the call.
13 . The system of claim 11 , wherein, when generating the machine likelihood score, the computer is further configured to apply an attack prediction model to the first set of background events and the second set of background events to generate the machine likelihood score over a first segment corresponding to the first audio signal and a second segment corresponding to the second audio signal.
14 . The system of claim 11 , wherein, when applying the event classifier on the first set of acoustic features, the computer is further configured to apply the event classifier to detect one or more first anomalous events within the first audio signal; and
wherein, when applying the event classifier on the second set of acoustic features, the computer is further configured to apply the event classifier to detect one or more second anomalous events within the second audio signal.
15 . The system of claim 11 , wherein the computer is further configured to:
apply an estimation model to the first set of acoustic features to determine a first quality metric indicating a first degree of change in quality through the first audio signal; and
apply the estimation model to the second set of acoustic features to determine a second quality metric indicating a second degree of change in quality through the second audio signal.
16 . The system of claim 15 , wherein the first quality metric and the second quality metric each include at least one of: (i) a respective noise type, (ii) a respective reverberation ratio, or (iii) a respective signal to noise (SNR) ratio.
17 . The system of claim 15 , wherein, when generating the machine likelihood score, the computer is further configured to generate the machine likelihood score based upon a comparison between the first quality metric and the second quality metric.
18 . The system of claim 11 , wherein the computer is further configured to:
extract the first set of acoustic features from a first background in the first audio signal; and
extract the second set of acoustic features from a second background in the second audio signal.
19 . The system of claim 11 , wherein the computer is further configured to identify the caller as a human caller in response to the computer determining that the machine likelihood score does not satisfy the similarity threshold.
20 . The system of claim 11 , wherein the computer is further configured to generate an indication, for a user interface, of the caller as a machine caller in response to determining that the machine likelihood score satisfies the threshold.