Bioanalog Acoustic Emotion Recognition by Genetic Feature Generation Based on Low-Level-Descriptors

Affective computing has grown an important field in today´s man-machine-interaction, and the acoustic speech signal is very popular as basis for an automatic classification at the moment. However, recognition performances reported today are mostly not sufficient for a real usage within working systems. Therefore we want to improve on this challenge by evolutionary programming. As a starting point we use prosodic, voice quality and articulatory feature contours. We next propose systematic derivation of functional by means of descriptive statistics. In order to analyze cross-feature information and feature permutations we use genetic algorithms, as a complete coverage of possible alterations is NP-hard. The final attribute set is at the same time optimized by reduction to the most relevant information in order to reduce complexity for the classifier and ensure real-time capability during extraction process. Classification is fulfilled by diverse machine learning methods for utmost discrimination power. We decided for two public databases, namely the Berlin emotional speech database, and the Danish emotional speech corpus for test-runs. These clearly show the high effectiveness of the suggested approach