Speech emotion detection using time dependent self organizing maps

Abstract

We propose a framework for speech emotion detection that maps acoustic features into high level descriptors that integrates time context. Our framework uses three different algorithms to integrate the temporal context. The first method is based on temporal averaging of the original features. The second algorithm derives the descriptors by clustering the data using self-organizing maps (SOMs) and computing the temporal average of the activity distribution of the original features on the map. The third algorithm uses multi resolution window analysis and SOMs to compute a 2-D map of emotions and derives high level trajectories representing the behavior of the original features on the map. Using a standard emotional database and K-nearest neighbors classifier, we show that the proposed framework is efficient for analysis, visualization and classification of emotions.