Toward Accurate Sleepiness Estimation Modeling from Speech: A Preliminary Study of 25-Hour Sleep Deprivation

Accurate estimation of sleepiness levels is crucial for managing fatigue risks. Access to speech signals could aid in determining sleepiness across different settings. Previous machine learning (ML) studies faced challenges in accurately estimating sleepiness depths, especially with varied sentence structures. Our goals were to propose a novel approach based on X-vector integrating sleepiness patterns from multiple sentences and identify optimal labeling strategies for ML model training in estimating sleepiness levels and detecting sleepiness. Speech data was recorded from a 25-hour sleep deprivation experiment with 2-hour intervals. Our ML models were trained using various labels including hours awake, perceived sleepiness, and reaction time. The best performance came from ML models using hours awake labels with our approach, achieving 0.81 accuracy, 0.82 sensitivity, and 0.80 specificity in detecting sleepy states. This study advances quantitative sleepiness estimation via speech signals, potentially mitigating fatigue risks in diverse environments.