UniVoice: Dual RAG for Real-Time Lecture Translation in Multicultural Higher Education

초록

This study reconceptualizes the lecture-comprehension difficulties experienced by international students in Korean higher education not as individual language deficiencies, but as structural inequalities in knowledge access arising through educational migration in a multicultural society. While international students may have physical access to classrooms, they often face barriers in understanding lectures, participating in discussions, completing assignments, and engaging in academic assessment. In real-time lecture environments, disciplinary knowledge is delivered rapidly and interaction opportunities are limited, making language barriers a constraint on both knowledge acquisition and academic participation. To analyze these challenges, the study integrates perspectives from educational migration and multicultural higher education with Bourdieu’s linguistic capital, Cummins’s distinction between basic interpersonal communication skills and cognitive academic language proficiency, Fairclough’s critical discourse analysis, Fraser’s parity of participation, Fricker’s epistemic injustice, and science and technology study approaches to sociotechnical mediation. Based on this framework, the study proposes UniVoice, a sociotechnical model for supporting knowledge access in multilingual higher education. UniVoice combines instructor-side microphone input, LiveKit-based media transmission, Azure Speech STT, Azure AI Search, OpenAI-powered translation, Azure TTS, and student-side audio and caption delivery. A key contribution is the proposed Dual Retrieval-Augmented Generation architecture, which integrates course-material retrieval and disciplinary-knowledge retrieval to supplement lecture context, technical terminology, and conceptual relationships often missed by conventional sentence-level translation systems. Rather than assuming that artificial intelligence (AI) translation automatically promotes educational equity, the study critically examines issues such as data bias, language-quality disparities, cultural distortion, privacy concerns, and accessibility limitations. It concludes by proposing design principles, evaluation criteria, and governance conditions for the responsible adoption of AI-based real-time lecture translation systems in multicultural higher education.

키워드