Experiment with Stress Detection in Phonation Signal Recorded in Open-Air MRI Device

The paper is focused at detection of stress level in the phonation by Gaussian mixture models (GMM) classification. The proposed method compares partial GMM recognition scores for normal speech represented by a neutral state and low-arousal emotions with positive valence and for stressed speech modelled by high-arousal emotions with negative pleasure. For creation and training of the GMMs of the normal and stress classes, the noise/emotional speech databases were used. Phonation signals of basic five vowels were recorded in states without any negative stimulation and after exposition to vibration and noise during scanning in the open-air magnetic resonance imager (MRI). The first experiments confirm the principal functionality of the developed system. A stable, measurable increase of a stress factor was observed in the phonation signals recorded after MRI scanning. Additional analysis shows a relatively high importance of the used number of mixtures and the database used for GMMs building. To obtain greater differences in detected normal and stress classes, a detailed analysis of used speech features must be performed.