Particle.news
Download on the App Store
3 ARTICLES
·
8h ago

Brief Training Sharpens Brain Signals for Deepfake Speech Without Boosting Detection

EEG identified distinct temporal markers after a 12-minute labeled session, highlighting a neural–behavioral gap relevant to future anti-fraud training.

While the conscious mind is often fooled by deepfakes, the auditory system begins to "tag" subtle acoustic differences between human and AI speech after brief training, even before the listener can reliably tell them apart. Credit: Neuroscience News

Overview

  • Thirty adults judged sentences as human or AI before and after training, with only minimal improvement in behavioral accuracy.
  • Temporal response function analysis showed greater separation between AI and human speech at roughly 55 ms, 210 ms, and 455 ms after training.
  • The training involved explicitly labeled examples of human and AI voices and lasted about 12 minutes.
  • Researchers from Tianjin University and the Chinese University of Hong Kong published the peer-reviewed study in eNeuro and reported no commercial conflicts.
  • The authors say these neural markers could guide longer or targeted training and aid detection strategies as text-to-speech systems become more lifelike.