Japanese zero-shot voice cloning, running entirely in your browser on WebGPU — no inference server. Text and reference audio never leave your device. Weights stream from the model repo on first run.
⚠ Use Chrome. This hosted page runs ~2–3× slower in Edge (same GPU, same model — Edge throttling, not user-fixable). Chrome is fast (~0.5× RTF, faster than real-time).
Requires a WebGPU browser (Safari needs the flag). First load downloads the ONNX weights (~0.65 GB fp16 / ~2.3 GB fp32) and is cached by the browser.