Improving Speech Emotion Recognition System Using Spectral and Prosodic Features

The detection of emotions from speech is a key aspect of all human behaviors, Speech Emotion Recognition (SER) plays an extensive role in a diverse range of applications, especially in human-computer communication. The main aim of this study is to build two Machine Learning (ML) models able to classify the input speech into several classes of emotions. In contrast, we extract a set of prosodic and spectral features from sound files and apply a feature selection method to improve the SER rate of the proposed system. Experiments are being done to evaluate the accuracy of the emotional speech system with the use of the RAVDESS database. We performed the efficiency of our models and compared them to the existing literature for SER. Our obtained results indicate that the proposed system based on Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) achieves a test accuracy of 69.67% and 65.04% respectively with 8 emotional states.