MiMo 2.5 TTS by XiaomiMiMo

Overview

MiMo 2.5 TTS is an audio generation model family from XiaomiMiMo for text-to-speech. It turns written text into spoken audio, while following natural-language direction about tone, pace, and emotion. You can use a preset studio voice, invent a new voice from a short description, or clone a voice from a reference recording. The output audio is ready for narration, characters, voice replies, and prototyping without manual editing.

What you can build

• Audiobook or article narration with controllable emotion and pacing
• Character dialogue for games and interactive fiction
• Voice responses for chat apps and assistants
• Voiceovers for product demos, tutorials, and explainer videos
• Brand voice prototyping with “voice design” descriptions
• Consent-based voice cloning for personalization and accessibility

Inputs

• The text you want the model to speak, entered as plain text. You can add a style tag at the very start in parentheses, like “(whispering)” or “(singing)”, to set the delivery.
• Optional inline audio tags inside the text, written in brackets, to control moment-by-moment delivery. Use them where you want pauses, breaths, laughs, sighs, or other vocal behaviors.
• An optional built-in voice selection for preset synthesis. Available preset voices include Mia, Chloe, Milo, Dean, 冰糖, 茉莉, 苏打, and 白桦.
• An optional 1–4 sentence voice description that invents a new voice. Describe age range, texture, mood, and speaking style in natural language.
• An optional voice sample file to clone a voice from a recording. MP3 and WAV work best; other formats are typically converted. Keep the sample clear and under 7 MB.
• Optional speaking instructions that guide performance (tone, emotion, pace, character). These instructions are not spoken aloud.
• Your preferred output audio format, as MP3 or WAV.

Outputs

• A single audio file in the format you selected (MP3 or WAV).
• Spoken delivery that reflects your style tag, inline audio tags, and any speaking instructions.
• If you provided a voice sample, the audio aims to match the reference speaker’s vocal identity while still following your performance directions.

Limitations

• The MiMo 2.5 TTS models are designed around an 8K-token request and 8K-token maximum output limit, so very long scripts may need splitting.
• Singing mode is tied to the preset-voice TTS model. Voice design and voice cloning modes don’t support singing.
• Voice cloning needs an authorized reference recording. Poor recordings (noise, music, heavy reverb, multiple speakers) reduce similarity and stability.
• Very long passages can benefit from segmenting and stitching, especially beyond roughly 2,500 Chinese characters.
• Misformatted style or audio tags may be read as literal text or may not trigger the intended effect.

Safety & compliance

• Only clone voices you own or have explicit permission to use. Don’t impersonate real people without consent.
• Treat voice samples as sensitive data. Don’t upload private recordings you can’t legally share.
• Don’t generate deceptive audio meant to defraud, mislead, or imitate protected identities.
• Follow all applicable laws, platform rules, and your organization’s consent and privacy policies when generating or distributing speech.