[en] Virtual reality has shown great potential in many fields, especially in business and psychology. By immersing someone in a new computer-generated reality, it is possible to create realistic, safe, and controllable simulations for research and training, as well as new three-dimensional-enriched consumer experiences and services. Most of these environments, especially in the metaverse, rely on virtual representations of people called avatars. The design and non-verbal behaviors of these avatars must be carefully crafted to provide a realistic and truly immersive experience. This paper aims to understand how avatar nonverbal behaviors (i.e., body posture, facial expression, and head movement) are perceived by users immersed in a virtual reality context, a very common situation encountered in many simulations and especially during training. Therefore, the first objective of this study is to validate, through an experiment with 125 participants, how the audience's levels of emotional valence and arousal are perceived in virtual reality. Based on these results, a library of audience non-verbal behaviors corresponding to different arousal and valence levels is now available for future applications. The experiment also examines the benefits of using low-end versus high-end virtual reality headsets, and photo-realistic versus cartoon avatars. The results have implications for the design of realistic, challenging, and interactive virtual audiences.