hibari
Copyright 2026 Maureese Williams

Licensed under the Apache License, Version 2.0 (see LICENSE).

This product includes software and data developed by third parties:

Kokoro-82M (https://huggingface.co/hexgrad/Kokoro-82M), by hexgrad, Apache-2.0.
  hibari's acoustic model was distilled from Kokoro-82M (voice af_heart): its training audio,
  durations, pitch and energy targets were synthesized by Kokoro-82M. The phoneme vocabulary
  (assets/vocab.json) is Kokoro's.

kokoro (https://github.com/hexgrad/kokoro), by hexgrad, Apache-2.0.
  kokoro_patched/ in the research repository is a copy of the kokoro inference package, modified to
  return the teacher's durations, F0 and energy for distillation. Not part of the hibari package.

Training text: agentlans/high-quality-english-sentences
  (https://huggingface.co/datasets/agentlans/high-quality-english-sentences). Not redistributed.

misaki (https://github.com/hexgrad/misaki), by hexgrad, Apache-2.0.
  assets/lexicon_us_gold.json.gz and assets/lexicon_us_silver.json.gz are misaki 0.9.4's
  us_gold.json and us_silver.json, unchanged apart from compression. hibari/g2p/lexicon.py and
  hibari/g2p/__init__.py port misaki's English G2P rules (en.py) to run without spaCy.

spaCy (https://github.com/explosion/spaCy), by ExplosionAI GmbH, MIT.
  assets/tokenizer_exceptions.json is spaCy 3.8's English tokenizer special-case table;
  hibari/g2p/tokenize.py follows spaCy's prefix/suffix/infix tokenization rules. The tagger
  (assets/tagger.json.gz) was trained on tags produced by spaCy's en_core_web_sm model.

espeak-ng (https://github.com/espeak-ng/espeak-ng), GPL-3.0.
  Not included. The out-of-vocabulary net (assets/oov.onnx) was trained on pronunciations that
  espeak-ng produced for a word list, so that unknown words are read the way Kokoro's reference
  pipeline reads them. No espeak-ng code or data files ship with hibari.

StyleTTS2 (https://github.com/yl4579/StyleTTS2), by Aaron (Yinghao) Li, MIT.
  styletts2_gan/ in the research repository (discriminators and losses used while training the
  acoustic model) is adapted from StyleTTS2 and keeps its MIT license text. Not part of the
  hibari package.

num2words (https://github.com/savoirfairelinux/num2words), LGPL-2.1.
  Not included. hibari/g2p/numbers.py is an independent implementation of the English number
  words whose output was checked against num2words for compatibility with misaki.
