[laurieg]

Unfortunately, a simple normalization of kanji --> hiragana throws away pronunciation information.

[popalchemist]

You could just as easily use the LLM to convert the kanji into phonemes.

[xyzhg]

You can't lose word boundaries and phonemes don't tell you which part of the word is emphasized.

[popalchemist]

Modern TTS engines use tokenizers to convert words to phonemes. See: https://github.com/FunAudioLLM/CosyVoice/issues/202