Sarashina2.2-TTS

Sarashina2.2-TTS is a Japanese text-to-speech model released by SB Intuitions, described in arXiv 2606.25369. It targets kanji polyphony, the problem of choosing correct readings for kanji that have multiple pronunciations, using data scaling and targeted data synthesis.

The work introduces the Joyo Kanji Yomi Benchmark to evaluate TTS pronunciation correctness, with 13,095 test samples reviewed by 35 native Japanese speakers through a three-stage verification process. Weights and inference code are released on Hugging Face at sbintuitions/sarashina2.2-tts. The exact model license was not stated in the sources reviewed; check the model card before commercial use.