Speech Emotion Recognition with Classical and Deep Learning Models: An Empirical Study on the RAVDESS Dataset

This study presents a complete and reproducible workflow for speech emotion recognition using the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS). The dataset contains 1,440 speech recordings labeled across eight emotional categories: angry, calm, disgust, fearful, happy, neutral, sad, and surprised. A systematic pipeline was developed consisting of automated label extraction from file metadata, feature extraction using 40-dimensional Mel-Frequency Cepstral Coefficients (MFCCs), data normalization, and an 80/20 train–test split.