xiaomi/mimo-tts
Xiaomi MiMo V2.5 text-to-speech. Speaks text with one of 9 built-in voices, clones a voice from a reference clip, or invents a new voice from a written description.
Model Inputs
Text to speak. A style tag may lead the text, e.g. '(happy)Hello there'; inline tags such as '(sigh)' are also supported. Tags are never spoken out loud.
Built-in voice. Mutually exclusive with reference and description
A .wav/.mp3 clip whose voice gets cloned (Optional). 10-30 seconds of clean speech works best. Mutually exclusive with voice and description
Drag & drop a file here, or click to select
Supported formats: WAV, MP3 (max 10MB)
A sentence that invents a new voice, e.g. 'man in his forties, deep and raspy, speaks slowly' (Optional). Every call invents a slightly different voice. Mutually exclusive with voice and reference
Plain-language style note, e.g. 'slow down, sound tired' (Optional). Never spoken out loud. Not allowed together with description, which already carries the style
Output audio format
Result
Pricing
xiaomi/
mimo-tts
Pricing for Synexa AI models works differently from other providers. Instead of being billed by time, you are billed by input and output, making pricing more predictable.
For example, generating 100 requests costs about $2.00.
Check out our docs for more information about how per-request pricing works on Synexa.
| Provider | Price ($) | Saving (%) |
|---|---|---|
| Synexa | $0.02 | - |
| replicate | $0.04 | 50.0% |
Readme
MiMo V2.5 TTS is Xiaomi's text-to-speech model family, exposed here as a single endpoint. Which mode runs is decided by what you pass:
| You pass | What happens |
|---|---|
| nothing | Speaks with the default voice Dean, a deep English male voice |
voice | Speaks with one of the 9 built-in voices |
reference | Clones the voice from your audio clip |
description | Invents a brand-new voice from your written description |
voice, reference and description are mutually exclusive — passing two is an error rather than a silently ignored argument.
Capabilities:
- 9 built-in voices covering Chinese and English, male and female
- Voice cloning from a 10-30 second .wav/.mp3 sample
- Voice design: describe a voice in one sentence and the model builds it
- Fine-grained style control through natural language or inline tags — emotion, speed, pauses, breathing, accents, dialects and role-play
- Singing mode via the
(唱歌)tag at the start of the text (built-in voices only) - 24 kHz mono output as wav or mp3
Built-in voices:
| Voice ID | Language | Gender | Notes |
|---|---|---|---|
mimo_default | follows cluster | - | 冰糖 in China, Mia elsewhere |
冰糖 | Chinese | female | Clear, standard female voice |
茉莉 | Chinese | female | Slightly lower than 冰糖, gentler |
苏打 | Chinese | male | Younger-sounding male voice |
白桦 | Chinese | male | Deeper male voice |
Mia | English | female | English female voice |
Chloe | English | female | English female voice |
Milo | English | male | English male voice |
Dean | English | male | The deepest of all voices |
Chinese voices can read English and vice versa, but the accent is noticeable.
Style control:
- Write a plain-language note in
instructions, e.g. "slow down, sound tired". It is never spoken out loud. - Or embed tags in
text: a leading(style)tag sets the overall delivery, and inline(tag)markers adjust individual phrases. Supported styles cover basic and complex emotions, tone, timbre, character voices, dialects and role-play. instructionscannot be combined withdescription, which already carries the style.
Limitations:
- Reference clips must be .wav or .mp3; other formats are rejected. Base64-encoded size is capped at 10 MB.
- A voice built from
descriptiondiffers slightly on every call. To reuse one, save a generated sample and pass it back asreference. - Singing mode works with built-in voices only, not with cloning or voice design.