kokoro-pi
Copyright (c) 2026 Zachary Reece
Licensed under the MIT License (see LICENSE).

This project optimises and serves models it does not ship. It builds on:

Kokoro-82M
  https://huggingface.co/hexgrad/Kokoro-82M
  Copyright (c) hexgrad. Licensed under the Apache License, Version 2.0.
  The speech model itself. kokoro-pi downloads the ONNX export at install time and
  rewrites it; the weights are unchanged apart from the documented int8 quantisation
  of the vocoder's convolutions.

kokoro-onnx
  https://github.com/thewh1teagle/kokoro-onnx
  Copyright (c) thewh1teagle. Licensed under the MIT License.
  Provides the ONNX export, the voice pack, and the tokenisation and phonemisation
  front end this service calls. Model files are downloaded from its releases.

ONNX Runtime
  https://github.com/microsoft/onnxruntime
  Copyright (c) Microsoft Corporation. Licensed under the MIT License.
  The inference engine. The five public C and C++ API headers under native/include are
  vendored unmodified from release 1.23 (API version 23) so the custom operators can be
  compiled without a source checkout or a matching pip package layout.

espeak-ng, via espeakng-loader and phonemizer-fork
  https://github.com/espeak-ng/espeak-ng (GPLv3)
  Used at runtime for phonemisation, through the dependencies above. It is invoked as
  a separate library and is not linked into or redistributed by this project.
