Currently, expressive virtual humans are used in psychological research, training, and psychotherapy. However, the behavior of these virtual humans is usually scripted and therefore cannot be modified freely at run time. To address this, we created a virtual audience with parameterized behavioral styles. This paper presents a parameterized audience model based on probabilistic models abstracted from the observation of real human audiences (n = 16). The audience's behavioral style is controlled by model parameters that define virtual humans' moods, attitudes, and personalities. Employing these parameters as predictors, the audience model significantly predicts audience behavior. To investigate if people can recognize the designed behavioral styles generated by this model, 12 audience styles were evaluated by two groups of participants. One group (n = 22) was asked to describe the virtual audience freely, and the other group (n = 22) was asked to rate the audiences on eight dimensions. The results indicated that people could recognize different audience attitudes and even perceive the different degrees of certain audience attitudes. In conclusion, the audience model can generate expressive behavior to show different attitudes by modulating model parameters.
W. Brinkman, M. Neerincx and B. van Riemsdijk, "An Expressive Virtual Audience with Flexible Behavioral Styles," in IEEE Transactions on Affective Computing.