Real-time Stress Detection Model and Voice Analysis: An Integrated VR-based Game for Training Public Speaking Skills

Abstract

This paper describes a work in progress Virtual Reality (VR) based serious game, integrated with an algorithmic classification model for detecting stress during public speaking in real-time by analysing the speaker's voice. The designed VR game offers real-time virtual social support/feedback for the training of public speaking skills. We developed a stress detection model that recognises the stress and an altered normal confident state based on 24 actors' voice expressions from the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS). Three different classifiers models were constructed for each actor by extracting and identifying overall significant voice features. The results show that random forest classification using features Amplitude Envelope (AE), Root-Mean-Square (RMS) and Mel-Frequency Cepstral Coefficients (MFCCs) provides high accuracy for detection of stress, while many more features can possibly be explored.