An Evaluation of Emotion Units and Feature Types for Real-Time Speech Emotion Recognition

Abstract

Emotion recognition from speech in real-time is an upcoming research topic and the consideration of real-time constraints concerns all aspects of the recognition system. We present here a comparison of units and feature types for speech emotion recognition. To our knowledge, a comprehensive comparison of many different units on several databases is still missing in the literature and we also discuss units with special emphasis on real-time processing, that is, we do not only consider accuracy but also speed and ease of calculation. For the feature types, we also use only features that can be extracted fully automatically in real-time and look at which types best characterise which emotion classes. Gained insights are used as validation of methodology for our online speech emotion recognition system EmoVoice.