Local Video Transcription Pipeline

A local pipeline that turns raw video into usable material: a transcript, speaker-labelled dialogue, broadcast-ready subtitles and an illustrated PDF report, without sending a single frame to a cloud service. Whisper handles speech-to-text with word-level confidence scores, EasyOCR and a vision language model read what is actually on screen, and NVIDIA NeMo separates the speakers.

The interesting part is the correction stage. It fuses the visual vocabulary picked up from slides and interface text with phonetic matching, so terms Whisper mishears get repaired from what the video shows rather than guessed from context alone. Low confidence in the audio model loosens the matching threshold, letting the visual evidence take over where it is most needed.

A React dashboard drives the FastAPI backend and shows each stage as it runs. Models are pulled once and selected to fit the available GPU, so the whole thing keeps working offline.