Man or Machine? Toronto Company Masters Deepfake Detection with Advanced Voice Analysis
Eyes may be the windows to the soul, but at Klick Labs, it's all about the voice.
Klick Labs, the research arm of Toronto-based life sciences technology firm Klick Health, has developed a groundbreaking method to distinguish between human and AI-generated voices. This innovation comes at a crucial time as deepfakes—AI-created video, audio clips, or photos that seem real—are proliferating rapidly, fueled by the release of sophisticated AI chatbots. Celebrities and political figures, including Taylor Swift, President Joe Biden, and even the Pope, have fallen victim to this technology.
The Growing Threat of Deepfakes
The rise of deepfakes is alarming. Europol predicts that by 2026, up to 90% of online content may be synthetically generated. The Canadian Security Intelligence Service has labelled this trend "a real threat to a Canadian future." However, Klick Labs' senior vice president of digital health research and development, Yan Fossat, remains hopeful that their new technology can mitigate this threat.
Innovative Inspiration from Sci-Fi
Fossat and his team drew inspiration from science fiction classics like "Terminator" and "Blade Runner." In "Terminator," dogs can detect human impersonators, while "Blade Runner" features the Voight-Kampff machine, which measures physiological responses to differentiate humans from replicants.
Klick Labs gathered audio from 49 individuals with diverse backgrounds and accents for their project. These recordings were then processed by a deepfake generator to create synthetic clips. By analyzing vocal biomarkers—features embedded in voices that reveal information about the speaker’s health or physiology—the team developed a method to identify deepfakes.
Key Vocal Biomarkers
Klick Labs has identified 12,000 vocal biomarkers. However, their current deepfake detection method relies on five key factors: the length and variation of speech, the rates of micropauses and macropauses, and the overall proportion of time spent speaking versus pausing. Micropauses are less than half a second, while macropauses last longer. These pauses naturally occur when someone speaks and takes a breath or searches for words, something machines typically don’t mimic.
High Success Rate with Challenges Ahead
Klick Labs' technique currently boasts an 80% success rate in identifying deepfakes. However, AI technology is evolving, making it increasingly difficult to distinguish between human and machine-generated voices. For example, OpenAI recently introduced a new deepfake voice capable of mimicking micro-breaths, enhancing its realism.
Future Prospects
Despite these challenges, Klick Labs' research remains valuable. The team can test for other biomarkers like heart rate to enhance deepfake detection. Klick Labs is conducting 16 studies on vocal biomarkers and diseases, which could further support their deepfake detection efforts. One notable study has used vocal biomarkers to diagnose diabetes with high accuracy, a research line that will continue with Humber River Hospital in Toronto.
Conclusion
Klick Labs' advancements in voice analysis technology promise improved deepfake detection and potential for significant medical applications. As Fossat notes, rapid AI development means that staying ahead of deepfakes requires continuous innovation and adaptation.