kokoro-g2p

These are the English text front-end models that voxr's Kokoro-82M pipeline reads at run time. They mirror the g2p/ folder voxr expects beside a Kokoro checkpoint:

<kokoro dir>/g2p/tagger/   tagger.safetensors, tagger.json
<kokoro dir>/g2p/oov/      config.json, model.safetensors, vocab.json

tagger/

This is spaCy's en_core_web_sm 3.8.0 part-of-speech tagger, converted to safetensors.

• Contents: the embedding, encoder and softmax weights.
• tagger.json: the feature attributes, the hash seeds and the label set.
• Use: misaki's pipeline needs the tags to choose between POS-dependent pronunciations.
• License: the weights derive from en_core_web_sm (MIT, Explosion AI).

oov/

This is a small Llama-style character model that predicts Kokoro phonemes for words missing from misaki's lexicons.

• Input: <bos> <dialect> graphemes <sep>.
• Output: phonemes until <eos>.
• Training data: misaki 0.9.4's US and GB gold and silver lexicons (Apache-2.0, https://github.com/hexgrad/misaki).
• Vocab: vocab.json holds the token ids. Phoneme symbols are Kokoro-82M's.

License

• Repository: Apache-2.0.
• tagger/: also subject to the MIT license of spaCy's en_core_web_sm.