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 generating liveness scores for audio in calls, comprising:
obtaining, by a computer, a raw audio signal from a calling device including a speech signal for a speaker;
identifying, by the computer, a plurality of non-speech segments and a plurality of speech segments in the raw audio signal by executing a voice activity detection engine for detecting instances speech or non-speech in the raw audio signal;
extracting, by the computer, a plurality of background acoustic features for the plurality of non-speech segments and a plurality of acoustic features for the plurality of speech segments by executing one or more neural network architecture on the plurality of non-speech segments and the plurality of speech segments identified in the raw audio signal;
determining, by the computer, a plurality of scores based on the raw audio signal, the plurality of scores comprising (i) a first score identifying background change generated based upon comparing the plurality of background acoustic features extracted for the plurality of non-speech segments in the raw audio signal, and (ii) a second score identifying passive liveness generated based upon the plurality of acoustic features extracted for the plurality of speech segments of the speech signal of the speaker;
applying, by the computer, a machine-learning architecture to the plurality of scores to generate a liveness score indicating a likelihood that the speaker is a human; and
generating, by the computer, a classification of the speaker as one of human or machine based on the liveness score for the speech signal.
2 . The method of claim 1 , further comprising:
receiving, by the computer, a feedback dataset identifying a second classification indicating the speaker as one of human or machine; and
re-training, by the computer, the machine-learning architecture using a comparison of the classification and the second classification of the feedback dataset.
3 . The method of claim 1 , wherein obtaining the raw audio signal includes:
obtaining, by the computer, a training dataset including (i) the raw audio signal and (ii) a label indicating one of machine or human for the raw audio signal; and
updating, by the computer, the machine-learning architecture based on a comparison between the classification from the machine-learning architecture and the label of the training dataset.
4 . The method of claim 1 , further comprising determining, by the computer, based on at least one of feedback or training data, a threshold to compare against the liveness score to generate the classification.
5 . The method of claim 1 , wherein determining the plurality of scores further comprises applying, by the computer, the machine-learning architecture to extract the plurality of features from the raw audio signal and to generate the plurality of scores based on the plurality of features.
6 . The method of claim 1 , wherein determining the plurality of scores includes determining, by the computer, the plurality of scores including a third score identifying repetition of speech within the speech signal.
7 . The method of claim 1 , wherein applying the machine-learning architecture includes applying, by the computer, a plurality of weights defined by the machine-learning architecture to the plurality of scores to generate the liveness score.
8 . The method of claim 1 , wherein applying the machine-learning architecture includes applying, by the computer, a scoring neural network of the machine-learning architecture to generate the liveness score.
9 . The method of claim 1 , wherein obtaining the raw audio signal includes identifying, by the computer, from the raw audio signal, the speech signal corresponding to passive speech of the speaker.
10 . The method of claim 1 , further comprising providing, by the computer, via an interface, an indication of the classification of the speaker as one of human or machine.
11 . A system for generating liveness scores for audio in calls, comprising:
a computer comprising one or more processors and configured to:
obtain, a raw audio signal from a calling device including a speech signal for a speaker;
identify a plurality of non-speech segments and a plurality of speech segments in the raw audio signal by executing a voice activity detection engine for detecting instances speech or non-speech in the raw audio signal;
extract a plurality of background acoustic features for the plurality of non-speech segments and a plurality of acoustic features for the plurality of speech segments by executing one or more neural network architecture on the plurality of non-speech segments and the plurality of speech segments identified in the raw audio signal;
determine a plurality of scores based on the raw audio signal, the plurality of scores comprising (i) a first score identifying background change generated based upon comparing the plurality of background acoustic features extracted for the plurality of non-speech segments in the raw audio signal according to a voice activity detection engine, and (ii) a second score identifying passive liveness generated based upon the plurality of acoustic features extracted for the plurality of speech segments of the speech signal of the speaker;
apply a machine-learning architecture to the plurality of scores to generate a liveness score indicating a likelihood that the speaker is a human; and
generate a classification of the speaker as one of human or machine based on the liveness score for the speech signal.
12 . The system of claim 11 , wherein the computer is further configured to:
receive a feedback dataset identifying a second classification indicating the speaker as one of human or machine; and
re-train the machine-learning architecture using a comparison of the classification and the second classification of the feedback dataset.
13 . The system of claim 11 , wherein, when obtaining the raw audio signal, the computer is further configured to obtain a training dataset including (i) the raw audio signal and (ii) a label indicating one of machine or human for the raw audio signal; and
wherein the computer is further configured to update the machine-learning architecture based on a comparison between the classification from the machine-learning architecture and the label of the training dataset.
14 . The system of claim 11 , wherein the computer is further configured to determine based on at least one of feedback or training data, a threshold to compare against the liveness score to generate the classification.
15 . The system of claim 11 , wherein, when determining the plurality of scores, the computer is further configured to apply the machine-learning architecture to extract the plurality of features from the raw audio signal and to generate the plurality of scores based on the plurality of features.
16 . The system of claim 11 , wherein, where determining the plurality of scores, the computer is further configured to determine the plurality of scores including a third score identifying repetition of speech within the speech signal.
17 . The system of claim 11 , wherein, when applying the machine-learning architecture, the computer is configured to apply a plurality of weights defined by the machine-learning architecture to the plurality of scores to generate the liveness score.
18 . The system of claim 11 , wherein, when applying the machine-learning architecture, the computer is further configured to apply a scoring neural network of the machine-learning architecture to generate the liveness score.
19 . The system of claim 11 , wherein, when obtaining the raw audio signal, the computer is further configured to identify, from the raw audio signal, the speech signal corresponding to passive speech of the speaker.
20 . The system of claim 11 , wherein the computer is further configured to provide via an interface, an indication of the classification of the speaker as one of human or machine.