Real-time subtitle translation · Open source · Fully local inference

Live EN⇄ZH subtitles in your browser — audio never leaves your machine

AudioScript listens to an English meeting, podcast or video and puts live bilingual subtitles on top of your screen. Speech recognition and translation run inside your browser (ONNX Runtime WebAssembly) — no account, no upload, no per-minute fee.

Open the web app — free → Download desktop app

Chrome / Edge on desktop. English ⇄ Chinese. No signup required.

Why this exists

Every “AI subtitle” tool I tried wanted my meeting audio on their server and a subscription. For a 1-hour call, that is both a privacy problem and an expensive one.

AudioScript does the opposite: Whisper (speech → text) and an opus-mt model (translation) are downloaded once and executed on your device. After that, translating costs nothing and works offline.

What you actually get

Browser version (beta) · no install

  • Microphone mode — speak, get subtitles
  • Tab-audio mode — capture a browser tab's sound (YouTube, Meet in a tab, webinars)
  • Audio-file mode — drop an audio file and get a transcript + translation
  • Picture-in-Picture — keep the subtitle window floating over other apps
Try it now →

Honest limitations

Read this before you upvote — it will save you time

  • Chrome / Edge desktop only. Firefox cannot capture tab audio at all (browser limitation, not a setting).
  • English ⇄ Chinese only right now.
  • No SRT/VTT export yet. On-screen subtitles + transcript view only.
  • Quality is below paid cloud services. Quantized Whisper base/small models, tuned for latency, not for perfect accuracy.
  • It cannot hear other desktop apps. The browser build captures mic or a browser tab; system-wide audio needs the desktop app.

How it works

Three steps, no setup

  1. Open the web app. The quantized models download once (85 MB for the fast Whisper-base, 180 MB for the more accurate Whisper-small) and are cached by your browser.
  2. Pick a source: microphone, browser tab, or an audio file.
  3. Subtitles appear as you speak. Switch the target language EN → ZH or ZH → EN.

Desktop app additionally captures system audio on Windows / macOS / Linux.

FAQ

Is it really free?
Yes. MIT-licensed, no account, no usage limit. The models run on your hardware, so there is no server bill to pass on.

Where does my audio go?
Nowhere. Recognition and translation happen in the browser tab. You can disconnect the network after the models are cached and it keeps working.

Why is the quality not as good as Google/DeepL?
Because it runs on a laptop-grade model. Small quantized Whisper trades accuracy for speed and privacy. If you need broadcast accuracy, use a cloud service.

Why not Firefox?
Firefox does not implement tab-audio capture via getDisplayMedia. That is a browser capability gap we cannot patch from the page.

Can I use it for meetings?
Yes if the meeting runs in a browser tab (Google Meet, Teams web). Native desktop apps (Zoom client) need the desktop build.

Source & credits

Built on AudioScript (MIT), with Whisper and opus-mt models served through R2 and executed by transformers.js + ONNX Runtime Web. The browser build is an independent mirror; all credit for the desktop app belongs to the upstream author.

View on GitHub →