The prediction of fatigue using speech as a biosignal

Abstract

Automatic systems for estimating operator fatigue have application in safety-critical environments. We develop and evaluate a system to detect fatigue from speech recordings collected from speakers kept awake over a 60-hour period. A binary classification system (fatigued/not-fatigued) based on time spent awake showed good discrimination, with 80 % unweighted accuracy using raw features, and 90 % with speaker-normalized features. We describe the data collection, feature analysis, machine learning and cross-validation used in the study. Results are promising for real-world applications in domains such as aerospace, transportation and mining where operators are in regular verbal communication as part of their normal working activities.