Audio Description — PoC

Upload a video. Real silence detection (ffmpeg) finds gaps in the dialogue long enough to narrate; Claude looks at a real frame from each gap and proposes a description; a real local text-to-speech engine renders the proposed line and its actual spoken length is checked against the real gap duration — not estimated by the model. A reviewer listens to, edits, confirms, or dismisses each one, then exports a real AD script.