humanlayer-to-video

One Markdown document becomes a video: the left pane speaks a distilled version with karaoke word-highlighting as the voice says each word; the right pane shows the verbatim source, the active section highlighted. Both auto-scroll — the right slowly by section, the left fast by word.

The work splits in two: you distill the document into segments.json; one script (pipeline.ts) does everything mechanical — TTS, word-timestamp STT, alignment, build, render.

Prerequisites

dg (Deepgram CLI, authenticated — check dg whoami), ffmpeg/ffprobe, bun, and bunx hyperframes (run bunx hyperframes doctor once).

Step 1 — Distill the document into segments.json

The only creative step. Write a file: