Kokoroi-rs, an open-source Chinese text-to-speech (TTS) engine, has released version 0.1.2 with a significant architecture overhaul. The core change is the switch from International Phonetic Alphabet (IPA) to Bopomofo (Zhuyin) as the grapheme-to-phoneme (G2P) engine, which is more aligned with Chinese phonology. Additionally, the model architecture has been upgraded from a single model to three tiers (S, M, L), allowing users to balance between performance and resource usage. This upgrade addresses common issues in Chinese TTS, such as unnatural prosody and foreign-accented speech, by leveraging a phoneme system that better captures tonal and phonetic nuances. For developers and researchers in NLP, this represents a practical innovation that could improve the quality of Chinese voice applications, from virtual assistants to accessibility tools. The project is available on GitHub, and the update is particularly relevant for those working on multilingual or Chinese-specific TTS systems.
Kokoroi-rs v0.1.2 introduces a major architecture upgrade for Chinese text-to-speech, replacing IPA with Bopomofo phonemes and offering three model sizes (S/M/L). This shift aims to improve naturalness and efficiency for Chinese TTS, making it a noteworthy development for NLP engineers working on Asian languages.