Xiaomi opens MiMo-V2-TTS speech model via API
Xiaomi has made its in-house MiMo-V2-TTS speech synthesis model available through its MiMo API platform, free for a limited time, with free-form natural-language control over emotion, dialect, character voices and singing.
Xiaomi announced that MiMo-V2-TTS, a large-scale text-to-speech model it developed in-house, is now available through the Xiaomi MiMo API open platform. Access is free for a limited time. The page is headlined as a release of three "MiMo-V2.6" models, but the text only describes the TTS model.
Xiaomi says the model runs on a proprietary audio tokenizer and a multi-codebook architecture that models speech and text jointly. It was pretrained on "hundreds of millions of hours" of speech data and then tuned with what the company calls multi-dimensional reinforcement learning.
The main pitch is control. Users describe the style they want in their own words instead of choosing from preset keywords, for example "lazy, just woke up, slightly husky." According to Xiaomi, the model can:
- shift tone and blend emotions gradually within one utterance
- insert sounds such as laughter, coughs, sighs, pauses and hesitation
- speak in several Chinese dialects and accents, including Northeastern Mandarin, Sichuan, Henan, Cantonese and a Taiwanese accent
- voice characters such as the Monkey King, Lin Daiyu and Iron Man
- sing with accurate pitch and rhythm
The model also reads formatting cues without extra markup. It stresses text written in ALL CAPS and gives repeated words ("no no no no no") a matching rhythm and emotion.
Xiaomi gave no benchmarks, latency figures, comparisons with other TTS systems, language coverage beyond the examples above, or pricing for after the free period. It also did not identify the other two models implied by the headline.