A systematic evaluation of acoustic representation and feature fusion strategies for Arabic speech emotion recognition

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-09-30 · DOI: 10.1007/s10772-026-10290-9

Аннотация

Arabic speech emotion recognition (SER) in low-resource corpora demands comprehensive evaluation for trustworthy and reproducible outcomes. In this work, we rigorously analyze the EYASE corpus through six controlled experiments using handcrafted features, pretrained representations, and fine-tuned speech models. The proposed hybrid pipeline combines wav2vec2 self-supervised embeddings, handcrafted Spectro-temporal statistics and pitch descriptors via a soft-voting extra trees XGBoost ensemble. It achieves 72.4% accuracy and 0.72 macro-F1 on a held-out test split and 70.17 ± 1.78% accuracy over five seeds, surpassing the alternatives we analyze, including a CNN-BiLSTM baseline (60.3%) and fine-tuned emotion2vec (64.8 ± 3.3%). Feature-block ablation exposes high redundancy, where no block is needed. SHAP, LIME and Grad-CAM analyses reveal factors influencing forecasts. Finally, statistical confirmation of the observed performance differences is provided by Welch’s t test with Holm correction. Overall, this work offers a reproducible and interpretable standard for low-resource Egyptian Arabic SER and provides insights into various speech representations.

Читать у издателя →

Похожие статьи