Speech Emotion Recognition for Public Speaking Training in Virtual Reality

Public speaking often triggers strong emotional responses that influence both the speaker's performance and the audience's reactions. By immersing users in realistic speaking contexts in front of a virtual audience, Virtual Reality (VR) constitutes an effective training solution for such skills. To enable the development of responsive virtual agents, this work-in-progress presents research aimed at designing a Speech Emotion Recognition (SER) system compatible with VR environments. The prediction of ten emotions is targeted using various machine learning approaches, relying solely on the speaker's speech signal. To train the models, a bilingual acted corpus was used. Thanks to a perceptual validation study, the corpus was annotated with the emotions human raters effectively perceived. The proposed methodology is presented along with preliminary results.