Automatic Speech Recognition systems errors for accident-prone sleepiness detection through voice
Résumé
Excessive Daytime Sleepiness (EDS), a symptom linked to chronic sleepiness, impacts everyday life and increases risks of work or road accidents of subjects affected by it. The detection of accident-prone EDS through voice benefits from its ease to be implemented in ecological conditions and to be sober in terms of data processing and costs. Contrary to previous works, this study focuses on long-term sleepiness detection through voice. Using the Multiple Sleep Latency Test corpus, we propose a feature selection pipeline inspired by clinical validation practices to classify accident-prone EDSas measured by a threshold of 15 on the Epworth Sleepiness Scale-based on vocal clues. We propose three different approaches based on the acoustic quality of voice, reading mistakes, and a whole new approach, relying on Automatic Speech Recognition systems errors. The classification system achieves performances on the same scale as the state-of-the-art systems on short-term sleepiness detection through voice (74.2% of Unweighted Average Recall). Moreover, we give insights into the decision process implied during classification and the system's specificity regarding the threshold delimiting the two classes Higher-risk driver and Lower-risk driver.