Multiple classifier applied on predicting microsleep from speech

The aim of this study is to apply a state-of-the-art speech emotion recognition engine on the detection of microsleep endangered sleepiness states. Current approaches in speech emotion recognition use low-level descriptors and functionals to compute brute-force feature sets. This paper describes a further enrichment of the temporal information, aggregating functionals and utilizing a broad pool of diverse elementary statistics and spectral descriptors. The resulting 45,088 features were applied to speech samples gained from a car simulator based sleep deprivation study. After a correlation-filter based feature subset selection, which was employed on the feature space in an attempt to maximize relevance, several classification models were trained. The best model (Support Vector Machine, dot kernel) achieved 86.1% recognition rate in predicting microsleep endangered sleepiness stages.