PHONON-2 / RUNS ON YOUR CPU

Transcribe audio on your device.

Turn English recordings from 0.1 to 15 seconds into text using Phonon-2. No GPU is needed.

The first run downloads about 177 MB of model files. Your browser can save them for later. Your audio is never uploaded.


If no file is selected, the English example recording will be used.

Experimental end-to-end version ready.

Your transcript will appear here.
What we built
  1. We used the existing Phonon-2 model. We did not train a new model.
  2. We split its data into 26 files. The numbers inside stayed the same.
  3. We wrote engine.js and a small WebAssembly file to run the model on your CPU.
  4. Your CPU does all the work. The model runs in a worker so the page stays responsive.
  5. Hugging Face serves the files. Your recordings stay on your device.
Phonon-2 · Fermion Research — Derived from NVIDIA Parakeet, CC BY 4.0. Attribution