Genetic Algorithm Based Feature Selection Applied on Predicting Microsleep from Speech

Within this study we apply a 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 an usually large feature set (45k) utilizing a broad pool of diverse elementary statistics and spectral descriptors. Several (un-) supervised subset selection methods including genetic algorithm based methods were employed on the feature space in an attempt to prune redundant dimensions. The resulting dimensionality reduced feature space was applied to speech samples gained from a car simulator based sleep deprivation study (N=12; 01.00-08.00 a.m.). Among the tested dimensionality reduction methods a simple correlation filter approach (130 features remaining) reached the best recognition rate (85.1%, SVM) in predicting microsleep endangered sleepiness stages.