Comparative Analysis of Spectrogram and MFCC Representations for Speech Emotion Recognition Using Machine Learning

Emotion recognition is a key area of research within human-computer interaction, addressing the growing need for systems that can respond to human emotional states. While advancements have been made, challenges remain, particularly in selecting appropriate datasets, identifying effective audio features, and optimizing classification models. This study explores how different audio feature representations, specifically Mel-Frequency Cepstral Coefficients (MFCC) and spectrograms, influence the accuracy of emotion classification. By extracting these features from the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and applying Random Forest (RF) and Support Vector Machine (SVM) classifiers, the research compares the performance of each feature-classifier pairing. Results indicate that RF and SVM classifiers with MFCC features achieved 50% accuracy, while spectrogram features led to 45% and 54% accuracy, respectively. These findings suggest that simpler models, when combined with appropriate features, can offer promising performance, contributing to more responsive and adaptive human-computer interaction applications.