Multilingual speech emotion recognition using IGRFXG – Ensemble feature selection approach

Abstract

In the field of Human–Computer Interaction, it is essential to recognize emotions through speech signals. Selecting the important features from speech signals is crucial to increase the accuracy of machine learning classifiers. Including unnecessary features diminishes model accuracy and increase system complexity. In this study, we propose a novel approach to tackle the feature selection challenge in speech emotion recognition. Our method employs a IGRFXG – ensemble feature selection approach, the name IGRFXG is derived from the three base feature selection techniques it combines: Information gain (IG), Random Forest (RF), and XGBoost (XG), which operates in two stages. In the first stage, to reduce the presence of unnecessary features, we propose an heterogeneous ensemble feature selection technique that integrates three distinct feature selection methods: Information gain, Random Forest and XG Boost. During this stage, we prioritize the features based on their importance score. In the second stage, we intersect the top features selected by various feature selectors and generate a feature kernel. This feature kernel is subsequently passed to machine learning classifiers. The proposed approach is evaluated on four publicly accessible datasets, namely RAVDESS, SUBESCO, EMOVO, and EMODB using six different machine learning classifiers. The performance of the ML models is assessed based on metrics such as accuracy, recall, precision, and F1 score. The performance assessments indicate that the maximum accuracies achieved are 79.28 %, 92.80 %, 95.00 %, and 90.96 % for the RAVDESS, SUBSECO, EMOVO, and EMODB datasets, respectively.