Xiaomi releases three TTS models, open-source ASR
Xiaomi has launched its MiMo-V2.5 speech series: three text-to-speech models you steer with plain-language instructions, plus an open-source recognition model it says leads on dialects, code-switching and noisy audio.
Xiaomi announced the MiMo-V2.5-TTS series and mimo-v2.5-asr. It calls the lineup a full speech pipeline for AI agents, covering both recognition and synthesis. The listing headline calls the release "MiMo-V2.6," but the announcement itself names every model as V2.5.
The three TTS models are on the Xiaomi MiMo Open Platform and are free for a limited time:
- mimo-v2.5-tts comes with several preset voices and lets users control speech rate, emotion and tone.
- mimo-v2.5-tts-voicedesign creates a new voice from a one-sentence description.
- mimo-v2.5-tts-voiceclone copies a target voice from a small number of samples.
All three follow style instructions written in natural language. Users can write a brief the way they would brief an actor, or use structured scripts split into character, scene and direction layers. Xiaomi says the layered format keeps a character's voice consistent in uses like audio dramas and game NPCs. The models also accept inline audio tags in Chinese or English, such as "(sudden pause)" or "(laughs softly)," which can be stacked within a passage. Xiaomi says they can also pick up rhythm, emotion and speaker traits from plain text with no prompt.
Mimo-v2.5-asr is open source. Xiaomi says it is "industry-leading" on Chinese-English bilingual speech, Chinese dialects, code-switching, heavy noise and multiple speakers. The material provided includes no benchmark figures, model sizes, licence terms or post-trial pricing, so these performance claims come from Xiaomi and have not been independently checked.