Adaptive multi-feature speech emotion recognition using improved whale optimization with Lévy-flight exploration and elite opposition learning
Systems and Soft Computing · 2026-12 · 9 · 200635
DOI: 10.1016/j.sasc.2026.200635 ↗Kurz gefasst
Fachartikel in der Zeitschrift Systems and Soft Computing, 2026-12. Verfasst von Wei Hou und 2 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Nanjing Xiaozhuang University.
Abstract (Originalsprache)
Speech emotion recognition (SER) supports affective human-computer interaction, yet practical systems remain constrained by redundant acoustic descriptors, limited cross-corpus transfer, and the computational demands of large pretrained representations. This study introduces an adaptive audio-only SER framework that combines complementary MFCC, Mel-spectrogram and first-order delta descriptors with a CNN-SE classifier and an improved binary Whale Optimization Algorithm (IWOA). The methodological novelty lies in a leakage-controlled feature-selection layer in which continuous WOA positions are diversified by Lévy-flight perturbations before sigmoid transfer, refined through Elite Opposition-Based Learning under the same binary fitness criterion, and optimized jointly for validation accuracy and feature compactness. The principal methodological contribution lies in a leakage-controlled framework that systematically integrates complementary optimization operators for feature selection from heterogeneous acoustic representations, rather than relying on any individual component in isolation. The learned mask is estimated inside speaker- or session-independent training folds and is then evaluated with multiple classifier backbones, separating feature-selector gains from architecture effects. Experiments on RAVDESS, CREMA-D, and IEMOCAP show that the full model obtains an overall accuracy of 92.34%, outperforming CNN and CNN+WOA baselines while reducing the selected feature ratio to 58%. Additional ablation, cross-dataset, confidence-interval, and efficiency analyses indicate that the improvement is associated with both the complementary acoustic representation and the enhanced feature-selection strategy. These findings demonstrate the potential of adaptive multi-feature optimization to improve the accuracy, compactness, and robustness of speech emotion recognition across the evaluated public benchmark datasets.
Förderung
- Nanjing Xiaozhuang University
Zitieren
APA
Hou, W., Miao, Y., & Xu, X. (2026). Adaptive multi-feature speech emotion recognition using improved whale optimization with Lévy-flight exploration and elite opposition learning. Systems and Soft Computing, 9, 200635. https://doi.org/10.1016/j.sasc.2026.200635
Vancouver
Hou W, Miao Y, Xu X. Adaptive multi-feature speech emotion recognition using improved whale optimization with Lévy-flight exploration and elite opposition learning. Systems and Soft Computing. 2026;9:200635. doi:10.1016/j.sasc.2026.200635
BibTeX
@article{Hou2026,
author = {Hou, Wei and Miao, Yu and Xu, Xiang},
title = {Adaptive multi-feature speech emotion recognition using improved whale optimization with Lévy-flight exploration and elite opposition learning},
journal = {Systems and Soft Computing},
year = {2026},
volume = {9},
pages = {200635},
doi = {10.1016/j.sasc.2026.200635}
}
EUVIMED