Turn-taking is a cornerstone of full-duplex interaction, yet current AI models apply a single norm regardless of the scenario. This limitation stems from training data: human-human speech corpora capture natural timing but lack role grounding or scenario-specific norms, while heuristic or prompted synthesis injects behaviors without human preference grounding.
To address this, researchers introduce DuplexGen, a framework that generates dialogues with scenario-adaptive turn-taking by calibrating LLM predictions against a small set of slot-level human preference annotations. The approach spans six cooperative and competitive tasks, where human turn-taking preferences differ systematically.
Results show that DuplexGen aligns substantially more closely with human preferences than uncalibrated prompting or training solely on generic human-human data. A full-duplex model trained on DuplexGen-generated data exhibits distinctive, human-preferred turn-taking behaviors.
The authors conclude that human calibration, not corpus scale or prompt design alone, is what enables scenario-specific turn-taking synthesis.