Dynamic Prompting Improves Turn-taking in Embodied Spoken Dialogue Systems

Published in ROMAN' 25, 2025

A humanoid-robot SDS improves turn-taking by dynamically adjusting prompts to an audio LLM based on real-time floor perception (who’s speaking/listening). The LLM’s output drives both speech and floor-state updates, letting the system signal its stance. In robot interviews, this dynamic prompting outperformed static prompts: more accurate floor transitions, better signaling, fewer interruptions, smoother dialogue

Download paper here

Recommended citation: Yifan Shen, Dingdong Liu, Xiaoyu Mo, Fugee Tsung, Xiaojuan Ma, Bertram E. Shi

Recommended citation: Yifan Shen, Dingdong Liu, Xiaoyu Mo, Fugee Tsung, Xiaojuan Ma, Bertram E. Shi
Download Paper