Evaluating models of speaker head nods for virtual agents

Virtual human research has often modeled nonverbal behaviors based on the findings of psychological research. In recent years, however, there have been growing efforts to use automated, data-driven approaches to find patterns of nonverbal behaviors in video corpora and even thereby discover new factors that have not been previously documented. However, there have been few studies that compare how the behaviors generated by different approaches are interpreted by people. In this paper, we present an evaluation study to compare the perception of nonverbal behaviors, more specifically head nods, generated by different approaches. Studies have shown that head nods serve a variety of communicative functions and that the head is in constant motion during speaking turns. To evaluate the different approaches of head nod generation, we asked human subjects to evaluate videos of a virtual agent displaying nods generated by a human, by a machine learning data-driven approach, or by a hand-crafted rule-based approach. Results show that there is a significant effect on the perception of head nods in terms of appropriate nod occurrence, especially between the data-driven approach and the rule-based approach.