VLCaption

Auto-generate subtitles for whatever is playing in VLC. Fully local — no cloud APIs. On Apple Silicon it uses NVIDIA Parakeet via parakeet-mlx: more accurate than whisper-large-v3 in English, ~10-60x realtime, and it doesn't hallucinate captions over music or silence. Whisper covers ~100 languages as fallback.

Use it

VLC extension (recommended): play a video, open VLC > Extensions > VLCaption - Auto Subtitles, click Generate Subtitles. Subtitles load automatically when ready. The server starts on demand and exits a couple of minutes after VLC quits.

Watch mode (macOS): zero clicks — subtitles whatever VLC plays:

vlcaption watch    # --model auto --interval 3 --overwrite --include-audio --embed

Long files are progressive: a quick pass around the playhead puts subtitles up within seconds, then the full pass replaces them (--no-progressive to disable, --quick-window to size the first pass).

Either way, captions are saved as movie.srt next to movie.mp4, so they auto-load on every future open.

Install

# From PyPI — with uv
uv tool install "vlcaption[mlx]"   # Apple Silicon ([mlx] = fast engines)
uv tool install vlcaption          # elsewhere

# ...or with pip / pipx
pip install "vlcaption[mlx]"
pipx install "vlcaption[mlx]"

vlcaption install                  # set up the VLC extension (optional)

# Or from source
git clone https://github.com/WT-MM/VLCaption.git && cd VLCaption && ./install.sh

Needs VLC, Python 3.11+, and on a Mac ffmpeg (brew install ffmpeg). Manual extension install: copy vlcaption/extension/vlcaption.lua into VLC's lua/extensions/ directory and restart VLC.

Platform support

Developed and tested end-to-end on macOS (Apple Silicon, VLC 3.0.21) only, so far.

Reports and fixes from other platforms welcome.

Models

Models download on first use. auto = CoreML/ANE Parakeet on Apple Silicon (~1.7x faster than MLX at much lower power), Whisper turbo elsewhere (CUDA float16 on NVIDIA, int8 on CPU). The engine chosen is logged per job. No torch/MPS path on purpose: MLX and the ANE beat the MPS Whisper ports and CTranslate2 has no Metal backend.

Embed captions into the video

Lossless remux (mp4/m4v/mov/mkv/webm) — the .srt becomes a subtitle track inside the file:

vlcaption embed movie.mp4               # writes movie.subbed.mp4
vlcaption embed movie.mp4 --replace     # in place (atomic)

Server API

The extension auto-starts the server (port 9839). Run it yourself as vlcaption serve with --host/--port/--device/--exit-with-vlc. Hyphenated aliases (vlcaption-watch, vlcaption-server, ...) also exist. Endpoints: POST /transcribe (file_path, model, language, auto_load), GET /progress, GET /health, POST /shutdown.

Development

uv sync --extra dev --extra mlx
make format && make static-checks && make test

Architecture notes and the design review behind this layout: docs/redesign-2026-07.md.

Troubleshooting

• Extension missing from menu — check the .lua is in the extensions dir; restart VLC. Menu is under the VLC application menu on macOS.
• Menu item needs two clicks to reopen — VLC 3.x bug (#27688).
• "FFmpeg is not installed or not in your PATH" — brew install ffmpeg; if it's installed but the extension-started server can't see it, re-run vlcaption-install (GUI apps don't get Homebrew's PATH).
• Subtitles generated but invisible — audio-only files have no video surface; check Subtitles > Subtitle Track on videos.
• Debug logs — VLC > Tools > Messages at verbosity 2, prefix [VLCaption]; server logs: run the launcher in a terminal.

Uninstall

rm ~/Library/Application\ Support/org.videolan.vlc/lua/extensions/vlcaption.lua
rm -rf ~/.config/vlcaption
uv tool uninstall vlcaption   # or delete the cloned repo

Models live in ~/.cache/huggingface/.

License

MIT