Enhancing Speech Emotion Recognition Through Augmentation and Model Selection

This study investigates how to improve Speech Emotion Recognition by employing data augmentation strategies and carefully choosing models to increase classification accuracy. The primary goal is to improve deep learning and machine learning models' capacity to accurately detect emotion from speech data by taking into account the variability in an enriched dataset. We made an effort to apply methods that leverage feature extraction through MFCCs, ZCR, and RMSE using the datasets RAVDESS, SAVEE, CREMA-D, and TESS. Finally, we examine CNNs, Random Forest, and Multi-Layer Perceptrons. Based on our studies, we demonstrate that CNN performs better with supplemented data, obtaining an accuracy of 85.47 %. The current research shows that adding noise, pitch shifting, and temporal stretching to data improves model resilience and is essential for obtaining dependable SER performance.