Xiaomi Unveils MiMo-V2-TTS: Giving Agents a Voice with Soul

In the emerging Agent era, intelligent agents need not only perception and action but also expressive vocal ability; Xiaomi therefore releases MiMo-V2-TTS, a speech‑synthesis large model designed for multimodal interaction.

Model overview : The model is named MiMo-V2-TTS and uses the self‑developed MiMo Audio Tokenizer together with a novel multi‑codebook speech‑modeling architecture that captures fine‑grained acoustic features.

Training pipeline : First, the model undergoes ultra‑large‑scale speech‑text mixed pre‑training on more than 100 million hours of diverse audio, learning cross‑modal alignment and generation. Afterwards, a small set of high‑quality supervised data fine‑tunes the model for multi‑granular, multi‑style instruction control. Multi‑dimensional reinforcement learning then optimizes natural prosody, stable audio quality, accurate phoneme rendering, high‑fidelity voice cloning, and scenario‑appropriate tone.

Generalizable style control : Users can issue natural‑language commands to set a global voice tone and also adjust local emotional nuances within a single sentence, enabling smooth tone shifts and emotional progression.

Robust text understanding : By leveraging massive text‑speech alignment data, the model automatically interprets punctuation, interjections, and emphasis markers, converting them into appropriate prosodic patterns without any extra annotation.

Beyond plain speech : MiMo-V2-TTS supports multiple dialects (e.g., Northeastern, Sichuan, Henan, Cantonese, Taiwanese), role‑play style rendering, and high‑quality singing synthesis, demonstrated with example lyrics that preserve pitch and rhythm.

Next steps : The roadmap includes extending coverage to more languages beyond Chinese and English and deeply integrating the TTS with the MiMo-V2-Omni multimodal model, so agents can see, understand, and narrate the world with a lifelike voice.