Can you distinguish an AI-generated voice from a real one?
A new study from Hany Farid, a UC Berkeley professor of electrical engineering and computer science and researcher at the CITRIS and the Banatao Institute, doctoral student Sarah Barrington and Emily Cooper, a professor in the UC Berkeley Herbert Wertheim School of Optometry & Vision Science, has found that people cannot consistently identify recordings of AI-generated voices.
Targeting identity and naturalness in voice recognition, the team’s study found that humans can only tell when a voice is fake 60 percent of the time, and when an AI-generated voice and a real voice are put side by side, only 20 percent of the time can people tell they’re not the same identity.
To address this, Farid and his team have developed DeepSpeak, a large-scale dataset of real and deepfake audiovisual footage intended to develop new and further refine current deepfake detection techniques. DeepSpeak improves on previous deepfake data sets by prioritizing consensual data collection, using technologically advanced tools and generating a diversity of types of deepfakes.