Using Gemini and LuaLaTeX to transcribe physics videos into PDF/UA-2 and ISO 32005 math-accessible PDFs

Abstract

We present an AI transcription workflow that uses freely available tools to transcribe instructional physics videos with substantial mathematical content into math-accessible PDFs. Guided by our carefully crafted system instructions, Google Gemini transcribes linked or uploaded videos into LaTeX output. Importantly, all board-written or screen-presented equations are transcribed into LaTeX math and incorporated into the verbal flow or presented in descriptive segments, which also describe graphs, diagrams, and pedagogically important actions and interactions. Gemini's LaTeX output is appended to a prewritten SE tagging-enabled preamble and compiled by a LuaLaTeX engine; the resulting PDF transcripts routinely compile without errors and pass PDF/UA-2 and ISO 32005 validation tests. We provide a balanced performance assessment using numerous examples from test transcriptions. The publicly available supplementary resources contain complete transcripts corresponding to all presented examples, and also contain our well-tested system instructions and preamble so that others can use, adapt, study, and further develop our transcription framework.