Audio deepfakes, a new generation of AI tools, allows anyone to generate realistic audio simulating a person’s voice, making it easy to spread hateful speech and harmful disinformation that could impact public health. With fewer clues than videos, audio deepfakes are harder to detect, creating even more real-world consequences. And with vast amounts of voice recordings shared over the internet every day, coupled with politicians and celebrities with millions of followers being targeted, the threat is massive.

Our objective was to provide a simple solution to a growing problem. The Audio Deepfake Detector detects fake audio and promotes accuracy of information. It will provide a sense of security and help build trust in industries by detecting deepfake audio and making it harder to spread false information. The hope is that the detector will make the world a safer place by allowing people to make informed decisions based on authentic audio.

The Audio Deepfake Detector is the world’s first software that uses vocal biomarkers to detect inauthentic voice clips to stop the spread of harmful false information. It analyzes the speech for vocal biomarkers of biological signals, consistent with a human delivering the message. While a deepfake might include some of these biomarkers, they will be irregular, intermittent, or inconsistent with a living human, thus detecting that the audio is inauthentic. Because the current standard in deepfake detection is to use AI, we’re caught in an arms race between two branches of AI: one for the generation and one for the detection of deepfakes. The Audio Deepfake Detector uses a different approach: harnessing vocal biomarkers- hidden physiological signals-in the human voice that will be missing in a deepfake voice.

The detector will analyze an audio file and identify biological and physiological characteristics from the voice. If these characteristics are not identified or are irrational, (due to a machine creating the audio and not a person), it will alert the user that the audio is deepfaked.

We used a mixture of statistical tools and mathematical formulations, brought together in a completely unique combination to create an audio deepfake detector. Audio files are rapidly processed to extract vocal features such as pitch, tone and more esoteric properties such as fourier distribution. Characteristics such as age, BMI, lung function, blood pressure, glucose levels, or hormone levels can be measured from these extracted voice features.

Some of the vocal biomarkers have been validated, published in peer reviewed journals and patented, while others are still in development. The ensemble model of biomarkers to deepfake should be piloted by late 2023.

(2019) Initial Voice Deepfake Detection: A preliminary synthetic voice detector was created using Machine Learning. This device had a real world accuracy of ~70% in distinguishing between audio from real humans and faked audio. (Clinical trial 1)

(2020 - 2022) Voice Biomarker Specialization: Through the use of advanced mathematics and data analysis, a system was developed that can accurately predict physiological markers from voice recordings. These markers include features such as age, body mass index, blood pressure, ovulation, and glucose levels.(Clinical trial 2)

(2023) Extensive Audio Deepfake Detection: Merging our experience in synthetic voice detection with our expertise in vocal biomarkers, we are devising a new technique for deepfake detection. Our plan is to make the Audio Deepfake Detector available as a google chrome plug-in, whatsapp audio screener and media vetting tool, using those mediums as a way to authenticate any audio recording distributed to avoid spreading misinformation.(Clinical trial 3)

With the vast number of recordings being spread over the internet each day, there are millions of opportunities to avoid the dangerous consequences of Audio deepfakes. Now with the Audio Deepfake Detector, already in Beta test version on select news outlets and soon to be available as a google chrome plug-in, whatsapp audio screener and media vetting tool, we have the power to authenticate any audio recording distributed to avoid spreading misinformation. The impact of this project will not only be felt in the media and health industries, but also in politics, finance, and many other areas of society where trust and authenticity are critical.