发表机构
Karlsruhe Institute of Technology; AppTek; Charles University; Sesame AI; University of Edinburgh（卡尔斯鲁厄理工学院; AppTek公司; 查理大学; Sesame AI公司; 爱丁堡大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出一种轻量级反馈语头，从全双工模型隐藏状态预测并触发反馈语，跨7B和1B模型规模泛化，生成更频繁且时机更佳的反馈语，质量与真实反馈语相当。
AI 中文摘要
反馈语，例如在对方可能仍在说话时发出的“嗯哼”等简短回应，是自然对话的核心，但全双工口语对话模型很少显式地对其进行建模。我们引入了一个轻量级的反馈语头，它从全双工模型自身的隐藏状态预测反馈语应何时开始。一旦该概率超过可调阈值，就会强制解码一个反馈语。将其附加到7B（PersonaPlex）和1B（F-Actor）模型上，它能够跨规模泛化。探测实验证实隐藏状态能预判真实人类的时机，生成评估显示反馈语更频繁且时机更佳。人类评分者认为生成的反馈语与真实反馈语相当。
英文摘要
Backchannels, brief acknowledgements like "uh-huh" produced while the other party may still be talking, are central to natural conversation, but full-duplex spoken dialogue models rarely model them explicitly. We introduce a lightweight backchannel head that predicts, from a full-duplex model's own hidden states, when a backchannel should begin. Once this probability crosses a tunable threshold, a backchannel is force-decoded. Attached to both a 7B (PersonaPlex) and a 1B (F-Actor) model, it generalizes across scale. Probing confirms the hidden states anticipate real human timing, and generation evaluation shows more frequent, better-timed backchannels. Human raters judge the resulting backchannels on par with real ones.
