EmotionNet: A Novel Hybrid Deep Learning Model for Arabic Speech Emotion Recognition

This study presents EmotionNet, a novel hybrid deep learning model designed for Arabic speech emotion recognition. EmotionNet integrates a Variational Auto-Encoder (VAE) for latent representation learning with a lightweight classification branch enhanced by latent-space refinement. Evaluated on the KEDAS dataset, which includes five emotionally acted categories, the model achieved a test accuracy of 93.99% and outperformed conventional classifiers such as SVM, MLP, and Random Forest. The proposed approach employs a compound loss function and KL annealing to jointly optimize reconstruction and classification. Although the results are promising, the acted nature of KEDAS may overstate real-world performance, highlighting the need for evaluation on spontaneous, multimodal datasets, an effort currently underway in an ongoing interdisciplinary project.