Xiaomi MiMo-V2-TTS

Beyond the Monotone

The agentic era is here. AI agents can see, hear, act, and call tools to solve real problems. But a truly intelligent partner shouldn't just execute — it should express.

MiMo-V2-TTS is built for this moment: full-modal interaction in the age of agents. It gives agents more than understanding — it gives them a voice with warmth, emotion, and soul. Not a passive text-to-speech engine, but a natural extension of how an agent communicates and connects:

• Contextual emotion awareness — picks up emotional cues from text and automatically matches the most natural tone and delivery
• Universal style adaptability — from formal announcements to casual conversation, the output stays natural and on-register
• Real-time, seamless interaction — keeps pace with the agent's reasoning, making conversation feel fluid and effortless

Introduction

Xiaomi MiMo-V2-TTS is Xiaomi's self-developed large-scale speech synthesis model. Built on a proprietary Audio Tokenizer and a multi-codebook joint speech-text modeling architecture, MiMo-V2-TTS is pretrained on over 100 million hours of speech data and further refined through multi-dimensional reinforcement learning.

What it delivers:

• Highly controllable, multi-granularity style control — from setting the overall tone of an utterance to fine-tuning local emotional nuances, including mid-sentence shifts in mood and gradual emotional transitions within a single phrase
• Natural prosody reproduction — faithfully reproduces the natural rhythm and cadence of human speech
• Singing capability — accurately captures pitch and rhythm when singing, without sounding artificial

How it's built:

• Large-scale speech-text joint pretraining — pretrained on over 100 million hours of speech data, establishing strong cross-modal alignment and unified understanding-generation capabilities
• High-quality supervised fine-tuning — fine-tuned with a small amount of high-quality supervised data, giving the model generalizable, arbitrary-granularity style control that responds to free-form instructions
• Multi-dimensional reinforcement learning — to further unlock the expressive potential accumulated during massive pretraining, we introduce multi-dimensional RL that balances stability with expressiveness. Beyond conventional supervised learning, MiMo-V2-TTS leverages RL during speech generation training to continuously optimize across multiple dimensions:
  • More natural prosody
  • More stable audio quality
  • More accurate word-level articulation
  • Higher-fidelity voice cloning
  • Contextually appropriate tone and delivery across diverse scenarios
• Multi-layer codebook architecture — thanks to the MiMo-Audio multi-layer codebook architecture, the model operates in a high-fidelity discrete token space that preserves the rich information present in the original speech signal. This allows speech-specific reward signals to directly guide optimization during the reinforcement learning phase, making multi-dimensional evaluation metrics more effective in shaping the generation process.

Natural Language Style Instruction and Voice Control

Flexible Text-Driven Customization of Delivery Style

Most TTS systems give you a dropdown menu of emotions: happy, sad, angry, neutral. Pick one. MiMo-V2-TTS gives you a text box instead.

Describe the voice you want in plain language — any language, any level of detail — and the model generates speech that matches your description. No predefined tags. No fixed vocabulary. Just natural language.

• "sleepy, just woke up, slightly hoarse"
• "cutesy baby voice, a bit whiny"
• "deeply affectionate, speaking slowly, almost whispering"
• "impassioned and emphatic, like giving a speech to a crowd"

The model also handles dialects and character personas with the same interface:

• Northeastern Mandarin, Sichuan dialect, Cantonese, Taiwanese Mandarin

• Character voices: Sun Wukong, Lin Daiyu

This isn't keyword matching — the model parses the semantic content of your style description and maps it to the corresponding acoustic features during generation. Compositional descriptions work naturally: "angry but trying to stay calm" produces a measurably different output than either "angry" or "calm" alone.

Fine-Grained Control of Acoustic and Non-Verbal Sound Events

Real human speech is not a clean sequence of words. It's full of coughs, sighs, hesitations, sharp intakes of breath, and nervous laughter — the paralinguistic events that carry as much meaning as the words themselves.

MiMo-V2-TTS generates these events as naturally integrated components of the speech output, not as audio clips spliced in after the fact. The model understands where these events belong in context and how they should sound given the surrounding text.

Supported events include:

• Coughing — natural throat-clearing and cough sounds
• Pauses — meaningful silences for deliberation or rhetorical emphasis
• Hesitation fillers — "um...", "uh...", reflecting real-time cognitive processing
• Sighing — exhalations and deep breaths conveying emotional state
• Laughter — from subtle chuckles to sudden bursts

Hear it in action:

Advanced Text Comprehension

Rich Text Understanding

When a human reads "You are UN-BE-LIEVABLE!" out loud, they don't pronounce each syllable the same way. The capitalization, the hyphens, the exclamation mark — all of these are instructions for how to speak. Most TTS systems ignore them.

MiMo-V2-TTS interprets typographic and formatting cues as prosodic signals:

• ALL CAPS → automatic stress and emphasis ("THIS IS IMPORTANT" sounds emphasized, not just louder)
• Character repetition → mapped to speech rhythm and emotional intensity ("不不不不不" becomes a rapid, emphatic refusal)
• Punctuation → shapes intonation contours (questions rise, exclamations punch, ellipses trail off)

Example inputs:

Inferring Speaking Style from Text Context

Perhaps the most consequential capability: MiMo-V2-TTS can infer the appropriate speaking style directly from the text content itself, without any explicit style prompt.

• A question is spoken with rising intonation — automatically.
• An angry outburst shifts to a sharp, clipped delivery — automatically.
• A tender confession softens in pace and volume — automatically.

This emerges from the language model backbone's deep semantic understanding. The model doesn't just convert text to speech — it reads the text, understands the emotional arc, and adapts its delivery sentence by sentence.

Singing Capability

MiMo-V2-TTS supports singing voice synthesis — within the same unified model that handles speech.

To our knowledge, this makes it the only commercially available TTS API that natively supports both speaking and singing generation. No separate model. No mode switching. The same architecture that delivers a whispered confession can belt out a pop chorus.

What's Next

MiMo-V2-TTS is a milestone in our voice technology roadmap, but it is not the destination.

On our roadmap: Expanded language coverage beyond Chinese and English. And tighter integration with MiMo-V2-Omni's multimodal understanding capabilities — enabling agents that not only see and understand the world, but speak about it with the full expressiveness of a human voice.

The voice agent era needs voices that are more than intelligible. It needs voices that are alive.

We're building them.