Stress classification based on speech features via convolutional neural networks

Abstract

A vast investigation has deduced the drawbacks of stress that have caused an impact on society. The observed adverse consequences have been found to contribute to a negative quality of life for individuals. If prolonged without seeking help, it results in severe physical and mental health problems such as post-traumatic stress disorder (PTSD), depression, stroke, insomnia, and others. The present way of detecting stress using the bio-signals method is considered an invasive and ineffective method of detecting patient stress. The local newspaper also reports on the rise in this issue, citing an increase in Malaysian suicide attempts as a result of various issues like the financial crisis, unemployment, and environmental issues. The Malaysian Ministry of Health (KKM) reported that 91.2% of over one hundred thousand inquiries to the psychosocial helpline required emotional support and counselling more than other help. As seen in previous studies, stress speech features provide better accuracy compared to stress images or video-based features. It is also more convenient since it is a non-invasive and contactless approach. It will not generate any unnecessary stress when measuring it since the patients are not required to wear any equipment during the stress measurement. This research study used speech features as the input data to classify stress into three levels (low stress, medium stress, and high stress). The use of three levels of stress classes assists health practitioners in properly producing a suitable remedy for patients. Mel-frequency cepstral coefficients (MFCCs) and Teager Energy Operator-MFCCs (TEO-MFCCs), which are made by combining MFCCs and TEO, are compared in this study. The classification is done using convolutional neural networks (CNN), a deep learning approach. The dataset was taken from an experimental study that was conducted on 50 students in tertiary education. This study’s combination of MFCCs and CNN has produced exceptional performance metrics with an overall accuracy of 95.67%, which beat the previous research that used an unscripted dataset (an experimental study) with 81.86%. Furthermore, our study demonstrated the highest performance outcomes, achieving 99.58% accuracy when utilizing the same proposed system in a scripted dataset. The contribution of this study includes: discovering suitable speech features to detect the presence of stress in speech; designing deep learning algorithms that offer higher accuracy in stress predictions; designing the stress experiment that can elicit the participants’ stress; and lastly, the unscripted dataset consisting of the speech produced by the participants during the stress experiment. Finally, it contributes to the current research field, and in further development, it will be an early stress detection tool using voice that will assist responsible parties in taking prompt action to provide further care to patients.