一句话：提出将学习者问题路由到合适 SE 专家角色的语音教育工具，实验显示可提升参与度与感知有用性，但短期记忆无增益。
生成式大语言模型已被软件工程（SE）学生广泛使用，但主流聊天界面给出的解释过于通用，与 SE 话语习惯和即时求助（just-in-time help-seeking）需求契合度低。作者设计并实现了一个语音到语音的教育工具，将学习者的问题路由到合适的 SE 专家角色（persona），并生成简洁、受角色条件约束的解释。为评估效果，他们开展了一项受控的、形成性被试内试点研究，24 名低年级 SE 学生分别使用该工具与公开的 ChatGPT 界面。结果显示，角色条件化系统在参与度（M=4.46 vs. 3.46）和感知有用性（M=4.13 vs. 3.75）上评分更高，感知理解提升较小（M=3.92 vs. 3.71）；短期事实回忆无显著差异（6.92/8 vs. 6.63/8），79% 的参与者更偏好该工具。作者认为交互框架设计能改善学习体验，但未必带来可测量的短期记忆收益，并据此提出开展纵向课堂研究以及面向负责任采纳的教师可配置控制机制。
Generative large language models (LLMs) are widely used by software engineering (SE) students, yet mainstream chat interfaces provide generic explanations that are weakly aligned with SE discourse and just-in-time help-seeking. We present a speech-to-speech educational tool that routes learner questions to an appropriate SE expert persona and generates concise, persona-conditioned explanations. In a controlled, formative within-subjects pilot study with 24 junior-level SE students, we compare the tool against the publicly available ChatGPT interface. The persona-conditioned system was rated higher for engagement (M=4.46 vs. 3.46) and perceived usefulness (M=4.13 vs. 3.75), with a smaller gain in perceived understanding (M=3.92 vs. 3.71). Short-term factual recall did not differ (6.92/8 vs. 6.63/8), and 79% of participants preferred the persona-conditioned tool. These findings suggest that interaction framing can improve learners' experience without measurable short-term recall gains, motivating longitudinal classroom studies and instructor-configurable controls for responsible adoption.