Improving a Speech Emotion Recognition System for Saudi Dialect

Speech Emotion Recognition (SER) is a challenging and essential task, particularly for underrepresented languages and dialects. This study introduces an improved SER system specifically designed for the Saudi dialect, addressing significant challenges such as data scarcity and imbalanced datasets. To tackle these issues, advanced data augmentation techniques—such as noise addition, volume adjustment, and time shifting—are applied to enhance dataset diversity and model robustness. The feature extraction process utilizes Mel-Frequency Cepstral Coefficients (MFCC), which effectively capture the key characteristics of speech signals. These features are then processed using traditional machine learning classifiers, including KNN, SVM, and LDA, as well as deep learning models such as CNN and MLP. The experimental results reveal significant improvements in accuracy, precision, recall, and F1-scores across all classifiers after implementing data augmentation. Among the classifiers, LDA and CNN achieve the highest accuracy of 96.19%, showcasing the efficacy of tailored data augmentation strategies in advancing SER for the Saudi dialect.