visotalk

Write a narrated explainer video as a document (doc.tsx), check it, and render it to MP4. The document holds the narration, the pictures it shows (parts of ordinary React) and the words each thing appears on; the engine owns time and rendering, and the same document renders to the same frames every time.

npx visotalk validate doc.tsx
npx visotalk render   doc.tsx      # → .visotalk/<id>/out/final.mp4

# or install once; the command is then `visotalk`
npm i -g visotalk
visotalk validate doc.tsx

The commands are listed in the header of the executable and in the agent skill that ships with VisoTalk (skills/visotalk/SKILL.md): validate, prepare, sheet, still, design, review, qa, render, play, schema.

Requirements

• Node.js 22 or later and ffmpeg on the PATH.
• Python 3.9 or later.
• Rendering uses a pinned chrome-headless-shell (Chrome for Testing).
• Narration uses a local Kokoro model, and word timing (English) a wav2vec2 aligner, both through Python.

validate and schema need none of these. visotalk doctor shows what is missing and visotalk doctor --fix installs it into ~/.visotalk (about 1 GB, checked against fixed sizes and SHA-256 sums); ffmpeg and Python are yours to install. To use your own copies instead, set VISOTALK_CHROME (a chrome-headless-shell binary), KOKORO_DIR (a folder with kokoro-v1.0.onnx and voices-v1.0.bin) and VISOTALK_PYTHON (a Python with kokoro-onnx, soundfile, torch and torchaudio). VISOTALK_HOME moves ~/.visotalk; keep that path short, because narration fails under one longer than about 90 characters.

License

MIT.