INTERSPEECH 2015
16th Annual Conference of the International Speech Communication Association

Dresden, Germany
September 6-10, 2015

Contemporary Stochastic Feature Selection Algorithms for Speech-Based Emotion Recognition

Maxim Sidorov (1), Christina Brester (2), Alexander Schmitt (1)

(1) Universität Ulm, Germany
(2) Siberian State Aerospace University, Russia

In this study a class of Multi-Objective Genetic Algorithms (MOGAs) is proposed to select the most relevant features for the problem of speech-based emotion recognition. The employed evolutionary algorithms are the Strength Pareto Evolutionary Algorithm (or SPEA), the Preference-Inspired CoEvolutionary Algorithm with goal vectors (or PICEA), and the Nondominated Sorting Genetic Algorithm II (or NSGA-II). Performances of the proposed algorithms were compared against conventional feature selection methods on a number of emotional speech corpora. The study revealed that for some of the corpora the proposed approach significantly outperforms the baseline feature selection methods up to 5.4% of relative difference.

Full Paper

Bibliographic reference.  Sidorov, Maxim / Brester, Christina / Schmitt, Alexander (2015): "Contemporary stochastic feature selection algorithms for speech-based emotion recognition", In INTERSPEECH-2015, 2699-2703.