University of Eastern Finland study: Humans worse than chance at detecting synthetic speech

A new study from the University of Eastern Finland’s Computational Speech Group reveals that humans are surprisingly poor at identifying deepfake voice recordings. In a localization task, 47 participants were asked to mark suspicious synthetic segments across authentic, fully synthetic, and partially synthetic utterances. The researchers manipulated three trust cues: instructional framing (how the task was introduced), affective priming (emotional framing), and provenance labeling (telling the origin of the audio). The results showed that utterance class was the primary determinant of detection accuracy—fully synthetic speech was detected at below-chance levels, meaning participants performed worse than if they had guessed randomly.

While trust cues produced no main effects on accuracy, they did motivate detection behavior, suggesting that people may change their search strategies but still fail to spot fakes. Interestingly, quality ratings on mechanicalness, expressiveness, clarity, and other dimensions tracked utterance type, indicating that participants could implicitly discriminate between real and synthetic speech even when they couldn't correctly identify it overtly. This has serious implications as AI voice cloning tools become widespread: even sophisticated human listeners cannot reliably detect deepfake audio, raising concerns for fraud, misinformation, and voice-based authentication systems.