Support Vector Regression for Automatic Recognition of Spontaneous Emotions in Speech

We present novel methods for estimating spontaneously expressed emotions in speech. Three continuous-valued emotion primitives are used to describe emotions, namely valence, activation, and dominance. For the estimation of these primitives, support vector machines (SVMs) are used in their application for regression (support vector regression, SVR). Feature selection and parameter optimization are studied. The data was recorded from 47 speakers in a German talk-show on TV. The results were compared to a rule-based fuzzy logic classifier and a fuzzy k-nearest neighbor classifier. SVR was found to give the best results and to be suited well for emotion estimation yielding small classification errors and high correlation between estimates and reference.