Study finds deepfake vocal tracts lack human anatomical correlation

Scientists have discovered that deepfake voices exhibit vocal tract configurations uncorrelated with human anatomy, even when synthetic audio remains perceptually indistinguishable from genuine speech. The research indicates that current generative models replicate acoustic output without simulating underlying biological constraints, creating detectable physiological artifacts invisible to human listeners. This anatomical divergence persists across multiple synthesis architectures, suggesting a fundamental limitation in how AI models generate vocalizations. The findings propose that forensic analysis of vocal tract geometry could serve as a robust detection mechanism independent of auditory perception. Researchers emphasize this biomarker exists regardless of audio fidelity improvements in future models. The study was highlighted by the Museum of Sound on October 7, 2026. This discovery potentially shifts detection strategies from signal processing artifacts toward biomechanical validation.