What gets downloaded
The first time you press Play with the natural voice, your browser fetches the ONNX version of the model and the voice files from Hugging Face, plus the ONNX runtime from a public CDN. That is roughly 90 MB on the WebAssembly path and more on WebGPU, and it is cached by the browser for next time. None of those requests include your text.
WebGPU or WebAssembly
If your browser supports WebGPU (recent Chrome and Edge on most desktops), the model runs on your graphics chip and is usually faster than real time. Otherwise it runs on the CPU with WebAssembly, which works everywhere but can be slower on older phones.
Using it offline
- Open ListenKit while online and play a sample so the model downloads.
- Later, open the page again without a connection; the app and model load from cache.
Browsers may evict caches when storage is low, in which case the model downloads again next time you are online.
FAQ
Is Kokoro free to use?
Yes. The model weights are released under the Apache-2.0 licence.
Does ListenKit send my text to a server?
No. Synthesis runs in a web worker on your device.