Vox
Speak a long, detailed instruction instead of typing it. Vox writes down what you said and shapes it into a clear prompt, ready to paste into your LM.
How it works
- Speak. Record a take. Whisper turns it into an editable transcript — fix names and jargon there.
- Shape. A small language model rewrites the transcript using a template: plain cleanup, a coding task, or a bug report. Templates are editable.
- Paste. Copy the prompt into whatever LM you work with.
The Image tab does the same for text in a screenshot or photo.
Private by construction
Every model runs in this browser, on this machine. Nothing you say, type, or paste is sent anywhere, and the page's security policy makes the browser refuse any connection except the one-time model downloads from Hugging Face.
First use
Models download once and stay in browser storage — after that Vox works offline and loads in seconds. Hover a light under Machine to see each model's size. Chrome or Edge with WebGPU gets the larger, faster models; elsewhere they run on the CPU.
The full README on GitHub — design decisions, benchmarks, and how to run your own copy.