Emotion Recognition based on Remarkable Change of Prosodic Features on Emotion Transition

Abstract

In this paper, we focus on emotion recognition using prosodic features in speech. It can be occurred that prosodic features on the utterances with two different emotions may show similarity. Therefore, distributions of each emotion are overlapping, so that the accuracy of emotion recognition goes down. This paper proposes the emotion recognition technique which can distinguish each emotion category in overlapped distributions. In dialogues, emotion changes cause the change of utterance expression. So it assumes that prosodic features change remarkably on the emotion transition. In proposed technique, to recognize emotion in overlapped distributions of emotion, we put remarkable changes of prosodic features on emotion transition to use. In the experiment, we treated 4 emotion transitions; "Interest to Sympathy", "Interest to Joy", "Interest to Doubt" and "Interest to Antipathy". We calculated the differences on prosodic feature between before and after the emotion transitions, and drew histograms of differences of prosodic features. We show that there are differences among emotion transitions in points of kinds of changed features and their amount of change. Moreover, we show that there are significant differences statistically among the amount of change of prosodic features.