Japanese LLM-TTS system and evaluation suite: a 0.8B model built on sarashina2.2-0.5b-instruct with CosyVoice/HiFT-GAN vocoding, trained on ~361K hours of Japanese and English speech with targeted augmentation covering all 2,136 Joyo kanji (weights shipped April 2026; paper June 2026). The paper introduces the Joyo Kanji Yomi Benchmark (2,136 kanji / 4,378 readings) and the Kana-CER metric measured with the newly released kana-whisper ASR (a 0.8B whisper-large-v3-turbo finetune to katakana, MIT) — attacking kanji-reading correctness, Japanese TTS's hardest failure mode.

Outputs 3

sarashina2.2-tts

model
Parameters 0.8B

Sarashina2.2-TTS Technical Report

paper

Joyo Kanji Yomi Benchmark

dataset
speechttsjapaneseopen-weight