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 of extracting fakeprints to evaluate risk of callers, comprising:
obtaining, by a computer, from a calling device, a raw audio signal including a speech signal for a speaker;
identifying, by the computer, using the raw audio signal, a plurality of datasets in accordance with a corresponding plurality of types of embeddings, the plurality of types of embeddings comprising a textual type of embedding and an audio type of embedding;
extracting, by the computer, executing a fakeprint extractor of a machine-learning architecture, a plurality of fakeprints for the speaker corresponding to the plurality of types of embeddings using the plurality of datasets, the plurality of fakeprints including a first fakeprint corresponding to the textual type of embedding and a second fakeprint corresponding to the audio type of embedding;
determining, by the computer executing a fakeprint evaluator of the machine- learning architecture, a risk score indicating a likelihood that the speaker is fake based on the plurality of fakeprints; and
generating, by the computer, a classification of the speaker as one of human or fake based on the risk score for the speech signal.
2 . The method of claim 1 , further comprising:
identifying, by the computer, a second plurality of fakeprints for the plurality of types of embeddings from a second raw audio signal associated with at least one of the calling device or the speaker;
comparing, by the computer, the first plurality of fakeprints with the second plurality of fakeprints to generate a similarity metric; and
wherein determining the risk score indicating the likelihood that the speaker is fake further comprises determining the risk score based on the similarity metric.
3 . The method of claim 2 , further comprising selecting, by the computer, a plurality of second raw audio signals associated with at least one of the calling device or the speaker received prior to the raw audio signal;
wherein identifying the second plurality of fakeprints further comprises identifying the second plurality of fakeprints for each of the plurality of second raw audio signals.
4 . The method of claim 1 , further comprising generating, by the computer, textual content from the speech signal for the speaker from the raw audio signal; and
wherein extracting the plurality of fakeprints further comprises extracting, using the textual content, at least one fakeprint including the first fakeprint of the plurality of fakeprints corresponding to the textual type of embedding of the plurality of types of embeddings.
5 . The method of claim 1 , further comprising detecting, by the computer, a plurality of temporal segments within the speech signal for the speaker from the raw audio signal, each of the plurality of temporal segments corresponding to a dialogue between the caller and an agent; and
wherein extracting the plurality of fakeprints further comprises extracting, using at least one of the plurality of temporal segments, at least one fakeprint of the plurality of fakeprints corresponding to a temporal type of embedding of the plurality of types of embeddings.
6 . The method of claim 1 , further comprising identifying, by the computer, metadata associated with the raw audio signal, and
extracting the plurality of fakeprints further comprises extracting, using the metadata associated with the raw audio signal, at least one fakeprint of the plurality of fakeprints corresponding to a metadata type of embedding of the plurality of types of embeddings.
7 . The method of claim 1 , further comprising storing, by the computer, an association between the plurality of fakeprints and at least one of the calling device or the speaker, the plurality of fakeprints to be compared against a second plurality of fakeprints extracted from a second raw audio signal from the calling device including a second speech signal of the speaker.
8 . The method of claim 1 , wherein determining the risk score further comprises determining a plurality of risk scores using the plurality of fakeprints, each of the plurality of risk scores corresponding to a respective type of embedding of the plurality of types of embeddings.
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 fake.
11 . A system for extracting fakeprints to evaluate risk of callers, comprising:
a computer comprising one or more processors configured to:
obtain, from a calling device, a raw audio signal including a speech signal for a speaker;
identify, using the raw audio signal, a plurality of datasets in accordance with a corresponding plurality of types of embeddings, the plurality of types of embeddings comprising a textual type of embedding and an audio type of embedding;
extract, executing a fakeprint extractor of a machine-learning architecture, a plurality of fakeprints for the speaker corresponding to the plurality of types of embeddings from the raw audio signal, the plurality of fakeprints including a first fakeprint corresponding to the textual type of embedding and a second fakeprint corresponding to the audio type of embedding;
determine, executing a fakeprint evaluator of the machine-learning architecture, a risk score indicating a likelihood that the speaker is fake based on the plurality of fakeprints; and
generate a classification of the speaker as one of human or fake based on the risk score for the speech signal.
12 . The system of claim 11 , wherein the computer is further configured to:
identify a second plurality of fakeprints for the plurality of types of embeddings from a second raw audio signal associated with at least one of the calling device or the speaker;
compare the first plurality of fakeprints with the second plurality of fakeprints to generate a similarity metric; and
determine the risk score indicating the likelihood that the speaker is fake based on the similarity metric.
13 . The system of claim 12 , wherein the computer is further configured to:
select a plurality of second raw audio signals associated with at least one of the calling device or the speaker received prior to the raw audio signal;
identify the second plurality of fakeprints for each of the plurality of second raw audio signals.
14 . The system of claim 11 , wherein the computer is further configured to:
generate textual content from the speech signal for the speaker from the raw audio signal; and
extract, using the textual content, at least one fakeprint including the first fakeprint of the plurality of fakeprints corresponding to the textual type of embedding of the plurality of types of embeddings.
15 . The system of claim 11 , wherein the computer is further configured to:
detect a plurality of temporal segments within the speech signal for the speaker from the raw audio signal, each of the plurality of temporal segments corresponding to a dialogue between the caller and an agent; and
extract, using at least one of the plurality of temporal segments, at least one fakeprint of the plurality of fakeprints corresponding to a temporal type of embedding of the plurality of types of embeddings.
16 . The system of claim 11 , wherein the computer is further configured to:
identify metadata associated with the raw audio signal, and
extract, using the metadata associated with the raw audio signal, at least one fakeprint of the plurality of fakeprints corresponding to a metadata type of embedding of the plurality of types of embeddings.
17 . The system of claim 11 , wherein the computer is further configured to store an association between the plurality of fakeprints and at least one of the calling device or the speaker, the plurality of fakeprints to be compared against a second plurality of fakeprints extracted from a second raw audio signal from the calling device including a second speech signal of the speaker.
18 . The system of claim 11 , wherein the computer is further configured to determine a plurality of risk scores using the plurality of fakeprints, each of the plurality of risk scores corresponding to a respective type of embedding of the plurality of types of embeddings.
19 . The system of claim 11 , wherein 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 fake.