Advancing Egyptian Arabic Speech Emotion Recognition: Insights from 2D Representations and Model Evaluations

This paper addresses the existing disparity in Speech Emotion Recognition (SER) advancements, where languages like English and German have made progress, while Arabic has received limited attention. The study aims to give real-time, effective techniques for identifying user emotions from speech, focusing on the widely spoken and comprehended Egyptian Arabic dialect. Human emotions play a crucial role in all of our interactions and relationships, determining expressions like happiness, despair, and grief. As technology integration grows, the ability of computers to perceive and comprehend human emotions becomes increasingly vital. Teaching computers emotional understanding enables personalized and adaptable systems. This study undertakes a comparative analysis of various audio image representations and classification models, utilizing EYASE [1] and EAED [20] datasets to identify the optimal ways to achieve this. After preprocessing the audios to eliminate noise, such as background music, the dataset undergoes augmentation to increase class samples. The Mel-frequency Cepstral Coefficients (MFCCs), Log-Mel spectrogram, discrete Meyer wavelet, and waveform audio features were used to represent each audio file. The models include a combination of VGG19 for features extraction & Support Vector Machine (SVM) for classification, ResNet50 with global average pooling and dense layer, and a custom Convolutional Neural Network (CNN). Notably, the CNN achieves the most promising outcomes, attaining accuracies of 90.1% for 3 emotion classes, 88.6% for 4 classes, and 76.3% for 5 classes through log-mel spectrograms.