Reflective, open-ended prompts are central to therapeutic conversations, yet conventional silence-based End-of-Turn (EoT) detection often misinterpret pauses as completion, cutting users off mid-thought. In our work, we explore the use of a pretrained large language model (LLM) as a foundation for EoT detection in therapeutic voice interactions by formulating prompts that encourage the LLM to evaluate whether a chain of thought still feels in progress or has reached a natural stopping point, we rely on the model’s semantic judgment rather than fixed silence thresholds. In controlled evaluations on therapy-inspired utterances, this LLM-based approach more reliably recognizes when users continue to formulate their thoughts compared to a silence-only baseline. Integrated into a prototype voice assistant and tested in a within-subjects user study, our system significantly reduces premature interruptions during reflective questioning and yields higher user ratings for conversational naturalness and empathy. This study serves as a proof of concept for alternative EoT mechanisms in the context of open-ended questions, with a particular focus on a qualitative evaluation of the perceived user experience. These results demonstrate that leveraging an off-the-shelf LLM, combined with pause cues, can support more patient-centered and empathetic turn-taking in therapeutic voice interfaces.