Abstract

The talk will show how models can be generated, which are capable of recognizing sequences of emotional states from speech. For this purpose Hidden Markov Models (HMMs) are introduced, which are trained on spontaneous, nonacted emotions. The user's emotion is tranferred into a 2-dimensional representation in the Valence- Arousal space using Plutchiks emotion wheel. Hence, not purely the basic emotions are recognized, but also an additional parameter, the word/syllable frequency, is extracted from the speech signal, which correlates with the user's arousal. Three models, two gender specific and a combined one will be presented and their results on unknown data will be discussed. For the evaluation of the robustness two cross-validation methods will be compared.