Combining classifiers with diverse feature sets for robust speaker independent emotion recognition

We consider two ways of combining classifiers for speaker independent emotion recognition: serial and parallel combination. In contrast to methods like bagging or boosting, our combination is based on different feature sets, having maximum diversity, instead of different training pattern sets. For that purpose, ensemble feature selection methods are presented for both combination types. For the parallel combination, we propose a novel method that has, to our knowledge, never been considered in the literature. The evaluation is performed on a well-known German emotional database [1]. Both new methods outperform the single stage and the hierarchical classifier presented in [2],[3] on the same database. Moreover, we examine the generalization capability of these classifiers when their feature subsets are not optimized directly on the test set. Here, the parallel combination proved to have the best generalization capability among all studied methods with a benefit of about 10%.