SPEECH EMOTION RECOGNITION USING SOME SELECTED MACHINE LEARNING TECHNIQUES

Abstract

This study addresses the challenge of Speech Emotion Recognition (SER) to enhance human computer interaction 
(HCI) by developing an interpretable, real-time system using the RAVDESS dataset, comprising 4,320 audio 
samples across eight emotions and two genders. The methodology employed Librosa to extract 20 Mel-Frequency 
Cepstral Coefficients (MFCCs) at a 22,050 Hz sampling rate, followed by training Support Vector Machine (SVM), 
Random Forest, and Gradient Boosting classifiers. Hyperparameters were optimized using 5-fold cross-validation 
and grid search on an 80-20 train-test split, with the system deployed via a Streamlit web application for practical 
evaluation. Results demonstrated that SVM achieved a test accuracy of 0.9206, a macro-averaged F1-score of 
0.9158, and completed training in 5 minutes, significantly outperforming Random Forest (0.7944) and Gradient 
Boosting (0.7493). MFCC coefficients 1-5, contributing 60% to classification, enhanced interpretability, while the 
web app processed predictions in 2 seconds with an 85% success rate across 50 test samples. Comparative analysis 
with five related works highlighted SVM’s efficiency and competitive accuracy among other machine learning 
models.