Emotion Recognition in Moroccan Arabic

This article explores speech emotion recognition (SER) for Moroccan Arabic (Darija) using deep learning. A dataset of 2,000 labeled audio samples across five emotions (happy, neutral, sad, angry, fearful) was collected from 80 speakers. Darija's linguistic diversity and regional variation pose challenges for emotion detection. After audio preprocessing and feature extraction (MFCCs, HNR, ZCR, F0, etc.), a ResNet-152 model was trained, achieving 90.12% accuracy. Angry and fearful emotions showed the highest recognition rates. This work highlights the potential of deep learning in under-resourced dialects and paves the way for emotion-aware Arabic applications in health, education, and virtual assistants.