No abstract

The problem of speech emotion recognition commonly is dealt with by delivering a huge feature set containing up to a few thousands different features. This can raise the "curse of dimensionality" problem and downgrade speech emotion classification process. In this paper we present minimal cross-correlation based formation of multi-level features for speech emotion classification. The feature set is initialized with most accurate feature and is expanded by selecting linearly independent features. This feature set formation technique was tested experimentally and compared with straightforward classification using predefined feature set. Results show superiority of our proposed technique by 5-25% for various emotion sets and classification settings.

The intensive research of speech emotion recognition introduced a huge collection of speech emotion features. Large feature sets complicate the speech emotion recognition task. Among various feature selection and transformation techniques for one-stage classification, multiple classifier systems were proposed. The main idea of multiple classifiers is to arrange the emotion classification process in stages. Besides parallel and serial cases, the hierarchical arrangement of multi-stage classification is most widely used for speech emotion recognition. In this paper, we present a sequential-forward-feature-selection-based multi-stage classification scheme. The Sequential Forward Selection (SFS) and Sequential Floating Forward Selection (SFFS) techniques were employed for every stage of the multi-stage classification scheme. Experimental testing of the proposed scheme was performed using the German and Lithuanian emotional speech datasets. Sequential-feature-selection-based multi-stage classification outperformed the single-stage scheme by 12–42 % for different emotion sets. The multi-stage scheme has shown higher robustness to the growth of emotion set. The decrease in recognition rate with the increase in emotion set for multi-stage scheme was lower by 10–20 % in comparison with the single-stage case. Differences in SFS and SFFS employment for feature selection were negligible.