What is Kokoro text-to-speech suitable for?
Kokoro is an 82 million parameter open-weight text-to-speech model. The tool loads the ONNX Community Kokoro-82M-v1.0 model with kokoro-js and runs inference inside a browser Web Worker, using WebGPU when available and falling back to WebAssembly.
The voice picker mirrors the model card catalog. American and British English voices are available in the current kokoro-js runtime; language packs listed by the model card remain visible as experimental options so the full catalog is easy to inspect. Generated WAV files are created on your device, not on The Browser AI servers.
Supported formats: WAV; single file limit: 500 characters.
How to use Kokoro text-to-speech
Step 1
Choose a voice
Select a language and voice, then adjust the speaking pace.
Step 2
Add your script
Use a model card example or enter up to 500 characters of your own text.
Step 3
Generate and download
The model downloads on first use, then creates a WAV you can preview and save.
FAQ
Is my text uploaded?
No. Text and generated audio stay in the current browser. The first run only downloads public model and voice files from Hugging Face.
Why are some languages marked experimental?
The Kokoro v1.0 model card lists several language packs, while kokoro-js 1.2.x currently exposes American and British English voices in the browser. Unavailable options are shown for catalog transparency.
Why is the first generation slower?
The browser downloads the lightweight 82M model and voice tensors on first use. Later generations can reuse the loaded model until the page is reloaded.
Technology and data sources
- Kokoro-82M-v1.0-ONNX — Model card, voice samples, language catalog, and ONNX weights on Hugging Face.
- Kokoro — The open-source Kokoro text-to-speech model and project documentation.
