Generating interaction gestures in dyadic conversations using a diffusion model.

Area of Science:

• Artificial Intelligence
• Computer Graphics
• Human-Computer Interaction

Background:

• Increasing expectations for computer graphic (CG) avatars and conversational robots necessitate improved dialogue skills.
• Current motion generation models often neglect crucial interaction factors between conversation partners.
• Natural conversational motion, like nodding, requires considering the behavior and voice of others.

Purpose of the Study:

• To develop an automatic interaction motion generation system for autonomous conversation systems.
• To enhance fluent human-computer interactions through multimodal dialogue skills.
• To generate natural, human-like interaction behaviors considering dyadic conversation dynamics.

Main Methods:

• Implemented an interaction diffusion model (IDM) utilizing a diffusion approach and masking features.
• The IDM accounts for two participants in dyadic conversations, using masks for conditional input generation.
• The model accommodates missing features and enables forecasting without retraining.

Main Results:

• The IDM successfully generates interaction behaviors for dyadic conversations.
• Experimental results indicate the model produces human-like behaviors during conversation.
• The generation of human-like behaviors is achieved within 30 milliseconds.

Conclusions:

• The developed interaction diffusion model (IDM) is effective for generating natural conversational motion.
• This approach significantly enhances the fluency and human-likeness of interactions with CG avatars and robots.
• The IDM offers a promising solution for advancing autonomous conversation systems.