Many second language (L2) learners struggle to speak English fluently despite years of formal study. Recent advances in generative AI (GenAI) voice chatbots offer promising opportunities to develop L2 oral proficiency. However, empirical evidence on their impact remains limited, and learners’ perceptions of their roles are underexplored. In this study we investigate how a voice chatbot (KELLY) supports L2 learners’ English oral proficiency development through a mixed-methods approach. KELLY is a GenAI voice chatbot specifically designed for engaging L2 learners in natural English conversations. We evaluate its impact and learners’ perceptions among 54 Chinese undergraduate students. The results revealed that the voice chatbot had varying effects on different aspects of oral proficiency: it significantly enhanced lexicogrammatical complexity, pronunciation accuracy, and speech fluency (p < .01), but had limited impact on lexicogrammatical accuracy. Learners’ feedback further attributed these effects to their engagement in meaningful communicative activities, receiving comprehension validation, and practicing in a non-judgmental environment. These findings highlight the potential and limitations of voice chatbots for L2 oral proficiency development and provide insights for future chatbot design and research.