Multimodal Distinctive Behavior for Expressive Embodied Conversational Agents

Abstract

Embodied Conversational Agents (ECAs) are a new type of computer interface which have human-like bodies and conversational skills. Users interacting with agents will be more engaged and participative if the agents exhibit behaviors which look coherent across different situations and emotional states. In the present work, we aim at increasing agents' believability by looking at two main aspects of the problem:
- an agent must be able to show its emotional state and communicative intentions not only through specific facial expressions, gestures, etc., but also by varying the quality of its movements (e.g., their speed, amplitude, etc.) and the choice of the modalities that are used to communicate;
- the agent must maintain a distinctive behavior, that is, its global behavior tendency has to remain apparent during any communication.
We have developed a model to solve the two above issues. We have evaluated the realism and believability of the resulting agent's behaviors through perceptual tests and application scenario. The resulting system is highly extensible and configurable. It can also be used as a research tool to study human communication.