tmoroney/auto-subs

AutoSubs – Offline AI Subtitles for Video Editors

What it does
AutoSubs is a desktop application that turns audio or video files into timestamped subtitle files — all on your own computer. It can run completely offline, supports dozens of transcription models (including OpenAI Whisper, Moonshine, Parakeet, SenseVoice, etc.), and can label speakers automatically. The generated subtitles can be exported as SRT or plain text, or sent directly into popular editing suites such as DaVinci Resolve, Adobe Premiere Pro, and After Effects.

Why it matters

• Privacy‑first – No cloud calls; everything stays on‑device, which is rare for high‑quality speech‑to‑text tools.
• Creative workflow integration – Direct plugins for the industry‑standard NLEs let editors generate and fine‑tune captions without leaving their timeline.
• Model flexibility – Users pick from a spectrum of local models, from tiny fast ones to large, high‑accuracy Whisper‑large or specialized models for Russian, Chinese, etc., letting them balance speed, RAM usage, and accuracy.
• Cross‑platform – Native installers for Windows, macOS (Intel & Apple Silicon), and Linux, plus Homebrew support.

How it works

1. Select a source file (audio or video) in the standalone UI or via the Resolve/Adobe extensions.
2. Choose a transcription model from the built‑in Model Manager; the app downloads the model files on demand.
3. The audio is fed to the chosen model (run through Whisper‑rs, ONNX Runtime, or other back‑ends). A VAD model isolates speech, a diarization model tags speakers, and optionally a forced‑alignment model adds word‑level timestamps.
4. The UI shows the generated subtitles for editing; you can adjust speaker labels, timing, or style.
5. Export – Save as an SRT file, copy to clipboard, or push the captions back into Resolve/Premiere/After Effects.

Key technical pieces

• Rust‑based core – Uses whisper-rs for Whisper, transcribe-rs for ONNX models, and pyannote‑rs for speaker diarization.
• Model manager – Handles on‑demand download, verification, and caching of model files (sizes range from ~60 MB to >3 GB).
• Integrations –
  • DaVinci Resolve: a Lua server + Fusion macro that lets the app read timeline audio and write caption tracks.
  • Adobe: a CEP extension that communicates with Premiere Pro/After Effects to import/export SRT.
• CLI – Full command‑line interface for scripting or headless use (see CLI.md).

Who should use it

• Video editors who need quick, accurate subtitles without paying for cloud services.
• Content creators concerned about data privacy.
• Developers who want a locally‑runnable speech‑to‑text stack and can extend the app via its open‑source Rust code.

Getting started
Download the appropriate installer from the Releases page (or brew install --cask auto-subs on macOS). After launching, follow the Quick‑Start steps in the README to generate subtitles, or explore the CLI guide for batch processing.

License & model terms
The application code is MIT‑licensed. Transcription models are downloaded separately and retain their original licenses (e.g., CC BY‑NC 4.0 for the MMS forced‑alignment weights). Users must comply with those model licenses.

---

All details are taken directly from the repository’s README; no additional features are inferred.