Cross-Corpus Classification of Realistic Emotions - Some Pilot Experiments
Florian Eyben, Anton M. Batliner, Björn Wolfgang Schuller, Dino Seppi, Stefan Steidl · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2010
We use four speech databases with realistic, non-prompted emotions, and a large state-of-the-art acoustic feature vector, for cross-corpus classifications, in turn employing three databases for training and the fourth for testing. Categorical and continuous (dimensional) anno-tation is mapped onto a representation of valence with the three classes positive, neutral, and negative. This cross-corpus classification is compared with within corpus classifications. We interpret performance and most important features. 1.