Omni Bridge is a Windows app that listens to whatever your computer is playing, a video, a call or a stream, or to your microphone, and shows a translation of it as captions in a small transparent window that stays on top of everything else. I started it in February.

Two halves

The window you see is a Flutter desktop app. The listening and translating happen in a Python server that the app starts on your own computer and talks to over a WebSocket, so the speech side can use Python’s audio and machine-learning libraries directly.

The server records the PC’s output through Windows’ loopback capture, so it hears exactly what you hear, with no extra hardware. A voice detector skips the silences, so nothing is sent to be recognised while nobody is speaking.

One job, several engines

Each caption passes through two steps: speech recognition turns sound into text, and translation turns that text into your language. Omni Bridge offers more than one engine for each, because each makes a different trade:

  • Speech recognition. Google’s online recogniser needs no setup. NVIDIA Riva is the most accurate across languages, and needs an API key. Whisper runs entirely on your computer, with no internet at all, in four sizes from Tiny to Medium that you download from Settings: the bigger the model, the more accurate and the slower.
  • Translation. Llama 3.1 8B is the one I recommend, because it takes the surrounding sentences into account. Google Translate covers over 100 languages with no setup. Google Cloud, MyMemory and NVIDIA Riva’s own translation are there too.

You can switch engines while it runs, without restarting. When the server connects, it first reports what your computer can do, including whether it has a graphics card and how much memory that card has, so the app knows which engines it can offer.

The overlay

The caption window can be dragged, resized, made more or less see-through, and collapsed to a single line. Every caption is also kept in a searchable history, so a line you missed is still there.

Free, with limits

Omni Bridge is free within a daily allowance of 5,000 tokens. Paid plans add engines and raise the limits, and anyone with their own NVIDIA API key can use the NVIDIA engines without limits on any plan.

Download it from omnibridge.marshalx.dev. The source is on GitHub.