VOICE-BASED PREDICTION OF DROWSINESS FOR ENHANCING AUTOMOTIVE SAFETY FEATURES

Abstract

This study proposes a novel, non-invasive method for predicting drowsiness based on short voice recordings, targeting applications in automotive safety and industrial risk management. Drowsiness remains a critical factor in traffic and workplace accidents, particularly in sectors where continuous alertness is vital. To address this issue, we developed a model that estimates the probability of heart rate decreases—a physiological marker associated with drowsiness—by extracting Mel-frequency cepstral coefficients (MFCCs) from speech and applying logistic regression analysis. Systematic data augmentation techniques were employed to enhance model generalization without requiring large-scale datasets. The proposed model achieved an Area Under the Curve (AUC) of 0.649 and an overall accuracy of 62.1%, demonstrating the feasibility of using voice as a proxy for physiological monitoring. Compared to conventional methods relying on EEG or HRV signals, our voice-only framework offers a passive, scalable, and practical solution for early-stage drowsiness detection. Future research will focus on incorporating prosodic and spectral-temporal features, adopting deep learning architectures, and integrating multimodal IoT-based sensing to improve robustness and real-world applicability.