Ensemble of SVM trees for multimodal emotion recognition

In this paper we address the sentence-level multi-modal emotion recognition problem. We formulate the emotion recognition task as a multi-category classification problem and propose an innovative solution based on the automatically generated ensemble of trees with binary support vector machines (SVM) classifiers in the tree nodes. We demonstrate the efficacy of our approach by performing four-way (anger, happiness, sadness, neutral) and five-way (including excitement) emotion recognition on the University of Southern California's Interactive Emotional Motion Capture (USC-IEMOCAP) corpus using combinations of acoustic features, lexical features extracted from automatic speech recognition (ASR) output and visual features extracted from facial markers traced by a motion capture system. The experiments show that the proposed ensemble of trees of binary SVM classifiers outperforms classical multi-way SVM classification with one-vs-one voting scheme and achieves state-of-the-art results for all feature combinations.