Hybrid Deep Learning Approach For Stress Detection Model Through Speech Signal

Abstract

Stress is a psychological condition that requires proper treatment due to its potential long-term effects on health and cognitive faculties. This is particularly pertinent when considering pre- and early-school-age children, where stress can yield a range of adverse effects. Furthermore, detection in children requires a particular approach different from adults because of their physical and cognitive limitations. Traditional approaches, such as psychological assessments or the measurement of biosignal parameters prove ineffective in this context. Speech is also one of the approaches used to detect stress without causing discomfort to the subject and does not require prerequisites for a certain level of cognitive ability. Therefore, this study introduced a hybrid deep learning approach using supervised and unsupervised learning in a stress detection model. The model predicted the stress state of the subject and provided positional data point analysis in the form of a cluster map to obtain information on the degree using CNN and GSOM algorithms. The results showed an average accuracy and F1 score of 94.7% and 95%, using the children's voice dataset. To compare with the state-of-the-art, model were tested with the open-source DAIC Woz dataset and obtained average accuracy and F1 scores of 89% and 88%. The cluster map generated by GSOM further underscored the discerning capability in identifying stress and quantifying the degree experienced by the subjects, based on their speech patterns