The eyes may be the windows to the soul, but at Klick Labs it's all about the voice.
The Toronto-based research division of life sciences and technology company Click Health has discovered a way to analyze voices in such detail that it can determine whether they are coming from a human or an artificially intelligent machine.
The feature comes as the emergence of several AI chatbots has led to an explosion in the number of deepfakes — AI-generated, lifelike videos, audio clips and photos — a phenomenon that has seen everyone from pop star Taylor Swift to US President Joe Biden and the Pope fall victim to the phenomenon.
And the trend isn't showing any signs of abating anytime soon: The European Union's law enforcement agency Europol recently predicted that by 2026, 90% of online content could be synthetically generated, a situation the Canadian Security Intelligence Service has called “a real threat to the future of Canada.”
But Jan Fossat, senior vice president of digital health research and development at Click Labs, hopes his company can help make the AI ​​world a little safer.
“Any unregulated technology is dangerous, and this one is moving faster than many others,” he said at his Toronto lab.
In a space cluttered with wires, appliances, and 3D printers, Fossat and his team of three began thinking about how their favorite sci-fi movies could help combat deepfakes.
“In The Terminator, they use dogs to sniff out whether the characters are human-like. In Blade Runner, there's a Voith-Kampff machine. I've always wanted to build a Voith-Kampff machine,” Fossat said, referring to the fictional test used in the film to measure physiological responses like eye movements and reaction times to determine whether characters are human or replicant.
For its own project, the Click team recruited 49 people with diverse backgrounds and accents, then fed their voices into a deepfake generator to create synthetic clips.
The clips were analyzed based on vocal biomarkers – features embedded in the voice that can tell us something about a speaker's health or physiology.
For example, if you've just run up the stairs, your breathing will be quicker and this will show in your voice, as will your voice if you've just woken up or are tired.
The Crick has identified 12,000 such biomarkers, but to distinguish humans from machines, Crick principal scientist Jaycee Kaufman said they've so far relied on five metrics: the length and variation in speech, the ratio of micropauses to macropauses, and the percentage of overall time spent in speech and pauses.
Micropauses last less than half a second, she says, while macropauses last longer. Micropauses often happen naturally when someone is speaking, such as taking a breath or searching for words.
“We don't pay much attention to it, but it's happening,” Fossat added.
“Humans have brains and need to think. They have lungs and need to breathe. Machines don't have that, so they don't do that.”
So far, Klick Labs' deepfake-detection method boasts an 80 percent success rate, but that success rate may not last long.
AI is constantly evolving and “getting better and better at making voices sound like humans,” making it increasingly difficult to tell if a clip is a deepfake, Fossat said.
“For example, OpenAI, the company behind (generative AI chatbot) ChatGPT, just released some breathtaking new deepfake voices a few weeks ago,” he said.
“It's really amazing to be able to replicate the microscopic breathing.”
He argues that this development does not make Click Labs' research useless, as there are thousands of other biomarkers, such as heart rate, that can be tested to detect deepfakes.
Klick Labs' 16 other studies into voice biomarkers and disease could also inform the company's research.
One study used voice biomarkers to diagnose diabetes, achieving 89% accuracy in women and 86% accuracy in men.
The research will soon continue with one that Click plans to conduct in collaboration with Humber River Hospital in Toronto, and Fossat said the study could eventually form the basis of a phone-based tool that anyone could use to find out their risk of contracting the disease.
Every advance in Crick's research means more opportunities to learn about biomarkers and apply them to disease and deepfake detection, which are proving difficult to address.
“Every time we do something things change so quickly that by the time we're done, everything has changed and we have to start again,” Fossat said.
This report by The Canadian Press was first published May 26, 2024.
