franken_whisper · live demo
Whisper in your browser, from the same Rust source
This page runs the franken_whisper native engine compiled to WebAssembly — the same
src/native_engine sources the CLI ships, one code path.
Your audio never leaves this device: the model downloads once (78 MB, cached and
SHA-256-verified in browser storage), then every transcription is local.
Nothing downloads until you click.
What this is (and is not)
- Model: Whisper tiny.en — English-only, the smallest tier. The native CLI runs large-v3-turbo; the browser build starts small on purpose.
- Speed: single-threaded wasm with SIMD128. The measured realtime ratio is shown after each run, on your hardware — no synthetic numbers. The native engine is much faster; this demo trades speed for zero install.
- Diarization: not in the browser yet. The Sortformer speaker model is 469 MB of f32 weights; shipping it here is a quantization project, tracked openly in the repo (bd-m2jm).
- Formats: mp3, m4a/aac, wav — decoded in wasm by Symphonia, resampled to 16 kHz mono.
- Privacy: no upload, no server, no analytics on this page. The only network request is the one-time model download from Hugging Face.