Feature Enriched Speech Emotion Recognition Using Mel Frequency Cepstral Coefficients

Accurate Speech Emotion Recognition (SER) continues to be a challenging task due to differences in speech patterns and interference from environmental noise. This paper tackles these challenges by introducing a deep learning-based solution that leverages audio data from the CREMA, RAVDESS, SAVEE, and TESS datasets. To process the data, this research extracts Mel Frequency Cepstral Coefficients (MFCCs) as key features and implement a Convolutional Neural Network (CNN) to analyze temporal and spectral patterns in the audio. The model incorporates dropout regularization to improve generalization and employs data augmentation techniques to enhance its robustness. Our method achieves consistent emotion classification, making a valuable contribution to the fields of human-computer interaction and affective computing.