The Power of a Nod and a Glance: Envelope vs. Emotional Feedback in Animated Conversational Agents

Justine Cassell and Kristinn R. Thórisson
Gesture and Narrative Language Group
M.I.T. Media Laboratory
20 Ames Street, E15-318
Cambridge, MA 02139
{justine, kris}

In this article we describe results from an experiment of user interaction with autonomous, human-like (humanoid) conversational agents. We hypothesize that for embodied conversational agents, non-verbal behaviors related to the process of conversation, what we call envelope feedback, is much more important than other feedback, such as emotional expression. We test this hypothesis by having subjects interact with three autonomous agents, all capable of full-duplex multimodal interaction: able to generate and recognize speech, intonation, facial displays, and gesture. Each agent, however, gave a different kind of feedback: {1} content-related only, {2} content + envelope feedback, and {3} content + emotional. Content-related feedback includes answering questions and executing commands; envelope feedback includes behaviors such as gaze, manual beat gesture, head movements; emotional feedback includes smiles and looks of puzzlement. Subjects’ evaluations of the system were collected with a questionnaire, and video tapes of their speech patterns and behaviors were scored according to {1} how often the users repeated themselves, {2} how often they hesitated, and {3} how often they got frustrated. The results confirm our hypothesis that envelope feedback is more important in interaction than emotional feedback, and that envelope feedback plays a crucial role in supporting the process of dialogue. A secondary result from this study shows that users give multimodal conversational humanoids very high ratings of lifelikeness and fluidity of interaction when the agents are capable of giving such