Spectral modulations (“timbre” – left panel) and temporal modulations (“linguistic rhythms” – right panel) characteristics of a sleepy voice. These markers of sleepiness have been derived through the explainability of AI, specifically Support Vector Machines, which were trained to recognize vocal samples from sleep-deprived individuals. This study underscores the importance of elucidating AI through interpretable acoustic features, shedding light on the operations of these powerful yet opaque systems.
Etienne Thoret, Thomas Andrillon, Caroline Gauriau, Damien Léger, and Daniel Pressnitzer.
Sleep Deprivation Detected by Voice Analysis
2024. PLOS Computational Biology 20 (2): e1011849 — @HAL