An Interaction Motion Generation Model Based on A Diffusion Model and Graph Attention for Back-Channel Behavior Synthesis

In this study, we propose a motion generation model based on graph attention networks and a diffusion probabilistic model for modeling back-channel gestures in dyadic conversation. Back-channel gestures, which include unconscious behaviors like nodding and body shifts, are a crucial component for achieving natural conversational agents. Our proposed method utilizes a graph attention network to mix the related information between the two participants’ behaviors. This approach explicitly handles the inter-feature interaction, thereby incorporating the influence of the partner’s behavior into a participant’s generated motion. We collected over 10 hours of dialogue data to train the proposed method and verify its motion generation performance. Experimental results showed that the use of a graph attention improved metrics such as the Fréchet Inception Distance. This suggests that the explicit consideration of the conversation partner’s behavior is important for modeling conversation behaviors.