Deepfakes – a term that, even on its quote-unquote “best” of days, is largely relegated to the likes of someone having put Tom Cruise’s face on Robert Downey Jr.’s to replace the iconic Iron Man look.

More often – on its worse days – deepfakes pose awfully significant problems.

Deepfakes, as we know, represent highly realistic manipulated assets (e.g. photos, videos, audio recordings) produced largely or entirely by AI. Those will ill intent typically conjure deepfakes to manipulate; to spread disinformation, wholly impersonate famous figures, damage reputations via nastily fabricated content. Basically, they undermine an already eroded trust in modern media. (Additional sinister examples like deepfake pornography, phony emergency broadcasts, etc.)

Two of the former examples – photo and video trickery – are often the most frequently pointed-at examples (and the threats they pose remain serious), but deepfake audio must be investigated further, too.

Though far from tone-perfect, recent incidents like fake President Biden robocalls and bogus Taylor Swift cookware ads have illustrated the growing sophistication of deepfake audio. Discerning a real voice from a synthetic one can be a challenge, even for the most trained ears. (Especially when it isn’t a globally renowned figure, but the cloned voice of a family member claiming immediate needs for cash or one’s boss demanding access to a platform. It’s problematic.)

This is why the researchers at Klick Labs ran a recent study aimed at fighting back against deepfake fraudsters.

Klick Labs – part of the Klick Group of companies that include Klick Health, Klick Katalyst, and Klick Consulting – conducted this study to focus primarily on “vocal biomarkers” present in human speech. These subtle, physiological indicators can present as changes in pitch, tone or speech rate, as well as breathing patterns, micropauses or tremors (to name a few), and placing nuanced focus on these aspects of vocal flow may allow others to more frequently detect deepfake technologies.

Klick’s researchers describe how vocal biomarkers, alongside machine learning, can be effectively used to distinguish deepfake audio from authentic audio with, quote, “reliable precision.” As part of its study, Klick Labs studied 49 participants from diverse backgrounds; varied ages, accents and so on. Counter-deepfake models were subsequently trained on voice samples the participants provided so the researchers could parse the authentic differences, then work on training AI models to do the same.

Some may see this usage of biomarkers as a near one-to-one for “fighting fire with fire.” Others may see this simply as “technology for good,” especially given an 80% accuracy rating for the models analyzing speech metrics and classifying the real versus the fake.

Yan Fossat, Senior Vice President of Klick Labs and the principal investigator of the study, says accuracy will pave the way to real progress.

“Our findings highlight the potential to use vocal biomarkers as a novel approach to flagging deepfakes because they lack the telltale signs of life inherent in authentic content,” Fossat explained. “These signs are usually undetectable to the human ear, but are now discernible thanks to machine learning and vocal biomarkers.”

Researcher need a promising path forward to fight against deepfakes, despite the ever-evolving nature of dynamic forgery technologies.

Learn more about Klick Labs here.