Cross-Lingual Lecture Note Generation for a Low-Resource Language: Design and User Evaluation of BaroPlatform AI for Somali-Speaking Students

Abstract

Somali is spoken by an estimated 25 million people yet remains among the least represented languages in speech and language technology. Somali-speaking university students consequently depend on English-medium lecture material while no AI-assisted learning tool produces study material in their language. This paper presents BaroPlatform AI, an end-to-end platform that converts English lecture video into structured Somali study material, and reports its evaluation with university students. The system ingests a YouTube link or an uploaded file, extracts audio with FFmpeg, transcribes it with OpenAI Whisper (whisper-1), and prompts Google Gemini (gemini-2.5-flash-lite, with gemini-2.5-flash as fallback) to generate Somali notes, a summary, key points, quiz items and transcript-grounded chatbot responses; Azure Cognitive Services synthesises Somali speech from the notes. The backend is FastAPI with PostgreSQL; the frontend is Next.js. Gemini was empirically rejected for transcription after it returned RECITATION safety blocks on widely distributed educational audio, an engineering finding reported here in full. Seven students at two Somali universities evaluated the deployed system using a 33-item instrument (Cronbach's α = 0.816). Usability scored 4.57/5, note quality 4.02/5 and educational value 4.45/5; every participant agreed without exception that the platform reduced the time needed to prepare study notes (M = 5.00, SD = 0.00), every participant reported measurable time savings, and every participant indicated they would use the system regularly if deployed. Operational data show that while 66.7% of submitted lectures completed the full pipeline, 100% of lectures that reached the generation stage produced usable Somali notes, locating the reliability constraint in media handling rather than in cross-lingual generation. Among the note-quality items, the naturalness of the generated Somali scored below the construct mean (M = 3.86), consistent with the known limits of general-purpose multilingual models on low-resource languages. We argue that prompt-level cross-lingual generation is a viable interim strategy for languages that lack the annotated corpora required for fine-tuning.