Scientists have learned to recognise audio deepfakes

Researchers at the University of Florida have developed a method for detecting audio deepfakes that measures differences between samples of organic and synthetic speech.

According to the researchers, today there are ways to reconstruct voice from the anatomical structure of the respiratory tract. With such methods, one could hear how even dinosaurs sounded.

To detect deepfakes, the researchers performed the reverse procedure. They modeled the vocal tract for organic and synthetic voices using audio recordings. This enabled them to reconstruct the anatomy of the speaker’s respiratory organs from a segment of the audio excerpt.

They found that deepfake recordings are not constrained by human vocal tract anatomy. In the modelling process, they observed forms that do not occur in humans.

According to the researchers, the method achieved 99% accuracy on a test data set.

The researchers said the study not only confirmed the hypotheses but also revealed other features. For example, for an audio deepfake the model often produced vocal tracts of the same relative diameter and consistency as a drinking straw. In reality, human organs are much broader and more complex in structure, the authors noted.

They say their approach will enable the detection of even audio deepfakes that sound convincing to the human ear.

“Fine, but biologically constrained aspects of speech generation by humans are not accounted for by modern models […]. Consequently, [anatomy] can serve as a powerful tool for detecting audio deepfakes,” the authors say.

Earlier in October 2021, it emerged that in the United Arab Emirates scammers forged the voice of the head of a major company and stole $35 million.

In the same month, researchers claimed the ability of audio deepfakes to deceive both speech-recognition devices and listeners.

Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!