A Hybrid Deep Learning Approach for Arabic Speech Emotion Recognition Using Acoustic and Spectrogram Features

Arabic Speech Emotion Recognition (ASER) presents unique challenges due to linguistic diversity, phonetic complexity, and the limited availability of labeled datasets. This work presents a hybrid deep learning model that integrates a Multi-Layer Perceptron (MLP) and a Convolutional Neural Network (CNN) to effectively classify emotions from Arabic speech data, using the KEDAS dataset. The model is designed to process two complementary sets of features: acoustic features, such as low-level descriptors, as well as Mel-spectrogram features that capture time-frequency information. Experimental results demonstrate that the hybrid architecture effectively leverages the strengths of both feature types, achieving a high accuracy of 98% in emotion recognition.