Selecting Training Data for Cross-Corpus Speech Emotion Recognition: Prototypicality vs. Generalization

We investigate strategies for selection of databases and instances for training cross-corpus emotion recognition systems, that is, systems that generalize across different labelling concepts, languages and interaction scenarios. We propose objective measures for prototypicality based on distances in a large space of brute-forced acoustic features and show their relation to the expected performance in cross-corpus testing. We perform extensive evaluation on eight commonly used corpora of emotional speech reaching from acted to fully natural emotion and limited phonetic content to conversational speech. In the result, selecting prototypical training instances by the proposed criterion can deliver a gain of up to 7.5 % unweighted accuracy in cross-corpus arousal recognition, and there is a correlation of .571 between the proposed prototypicality measure of databases and the expected unweighted accuracy in cross-corpus testing by Support Vector Machines.