A Hybrid System for Detection of Stress Using Human Emotions Through Voice

The art of detecting human emotions through voice has been here for quite some time now, but the last decade has shown a tremendous rise in different kinds of computer-based, automated assessment techniques. Affective computing is a field of artificial intelligence research and development concerned with developing systems and devices that can recognize, understand and process human sensibility. This paper describes noninvasive, robust and simple emotion detection and classification architecture based on machine learning techniques. There have been many research works done using image processing but machine learning approach for classifying emotion through voice signal is relatively new. The work present in this paper investigates the use of three types of machine learning model. The authors have been described an approach for emotion classification to detect stress. In the described approach, the emotion dataset has been classified with two different classifiers based on acoustic features. To improve performance of classifying algorithms principal component analysis (PCA) has been used. PCA was used for features selection. We used two different classifiers including the k-NN (nearest neighbour, kNN) and support vector machine (SVM) for the classification. The described methodology could be used for human emotion classification. As the first classifier, we have used the support vector machine classifier (SVM). The classification accuracy of 86.7% has been attained with complete dataset using SVM classifier. The average accuracy achieved 88.6% for classification with combination of first approach and SVM classifier has been obtained. The average accuracy has been achieved 89.2% with second approach and SVM classifier with PCA. The second classifier, we have used the k-NN classifier and has been obtained 85% classification. The average accuracy has achieved 90.1% for classification with first approach and k-NN classifier with PCA. The average accuracy has achieved 90.7% for classification using second approach and k-NN classifier with PCA. The author’s second strategy, which combines k-NN and PCA to classify stressed voice, has been recommended as the best accuracy at 90.7%.