Group-Wise Layer Aggregation of Self-Supervised Representations for Speech Emotion Recognition
IEEE Signal Processing Letters · 2026 · 33 · 2705-2709
DOI: 10.1109/lsp.2026.3707494 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Sunchan Park und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Gefördert von: Basic Science Research Program through the National Research Foundation, Ministry of Education, Advanced GPU Utilization Support Program.
Abstract (Originalsprache)
Self-supervised learning (SSL) models provide robust representations for speech emotion recognition (SER). However, conventional weighted-sum aggregation of all transformer layers can dilute critical emotional information. We propose Group-wise Layer Aggregation (GLA), which groups SSL layers based on their representational similarity using Centered Kernel Alignment (CKA) and processes each group through independent attentive pooling branches. This approach preserves the distinct cues encoded at different depths, preventing the information loss inherent in global aggregation. Experiments on IEMOCAP (76.4% weighted, 77.2% unweighted accuracy) and MELD (51.0% weighted F1-score) demonstrate that GLA outperforms recent approaches without auxiliary supervision. Ablation studies confirm that both similarity-based grouping and multi-branch pooling are essential components, validating the importance of leveraging the hierarchical structure of SSL models for emotion recognition.
Förderung
- Basic Science Research Program through the National Research Foundation
- Ministry of Education
- Advanced GPU Utilization Support Program
- Government of the Republic of Korea
Zitieren
APA
Park, S., Kim, H. S., & Kong, K. (2026). Group-Wise Layer Aggregation of Self-Supervised Representations for Speech Emotion Recognition. IEEE Signal Processing Letters, 33, 2705-2709. https://doi.org/10.1109/lsp.2026.3707494
Vancouver
Park S, Kim HS, Kong K. Group-Wise Layer Aggregation of Self-Supervised Representations for Speech Emotion Recognition. IEEE Signal Processing Letters. 2026;33:2705-2709. doi:10.1109/lsp.2026.3707494
BibTeX
@article{Park2026,
author = {Park, Sunchan and Kim, Hyung Soon and Kong, Kyeongbo},
title = {Group-Wise Layer Aggregation of Self-Supervised Representations for Speech Emotion Recognition},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2705--2709},
doi = {10.1109/lsp.2026.3707494}
}
EUVIMED