Recognition of emotional states in natural speech

Research in emotional speech recognition is generally focused on analysis of a set of primary emotions. However, it is clear that spontaneous speech, which is more intricate comparing to acted out utterances, carries information about emotional complexity or degree of their intensity. This research refers to the theory of Robert Plutchik, who suggested the existence of eight primary emotions. All other states are derivatives and occur as combinations, mixtures or compounds of the primary emotions. In this paper authors presents a novel approach for automatic speech clustering into subgroups representing primary emotions intensities. For this purpose a multimodal classifier based on k-means algorithm was implemented and tested on Polish spontaneous speech database. Studies have been conducted using prosodic features and perceptual coefficients. Results have shown that the proposed measure is effective in recognition of intensity of the predicted emotion.