This project does not use transformer.js and uses its own tokenizer. It uses easpeak-ng and onnxruntime-web packages delivered through cdns. It does not need node/npm to run.

a mainline npm package is available through https://github.com/hexgrad/kokoro

to use this project open a basic http server with index.html in the root directory.

demo on https://shubinwang.com/tts

to test the caching features https is required. model splitting feature is now optional since quantized model is so small.

running the model with multiple cores require the right server CORS headers

The orignal model by (kokoro-82m)[https://huggingface.co/hexgrad/Kokoro-82M/tree/main] by Hexgrad. The model in this project is quantized version by Xenova Joshua Lochner.

To use the original model(kokoro-v0_19.onnx) change modelVersion = 0 in index.js and make sure to refrence the right model location/name!

the model uses apache 2.0 lisence