MiMo Speech Synthesis

Use the MiMo-V2.5-TTS series through the ModelSell OpenAI-compatible Chat Completions endpoint for built-in voices, voice design, and voice cloning.

The MiMo-V2.5-TTS series expresses the target speech and voice instructions as Chat Completions messages. Generated audio is returned as Base64.

Endpoint

POST /v1/chat/completions

Authenticate every request with your ModelSell API key:

Authorization: Bearer $MODELSELL_API_KEY
Content-Type: application/json

Do not use the text-to-speech path

The MiMo-V2.5-TTS series does not use /v1/audio/speech. Always submit JSON to /v1/chat/completions and read the Base64 audio from the Chat Completions response.

Supported Models

Model	Use case	`audio.voice`
`mimo-v2.5-tts`	Synthesize with a built-in voice	Optional; defaults to `mimo_default`
`mimo-v2.5-tts-voicedesign`	Design a voice from a text description	Not supported; describe the voice in a `user` message
`mimo-v2.5-tts-voiceclone`	Clone a voice from an audio sample	Required; pass Base64 from an MP3 or WAV sample

Message Rules

Put the target text to speak in an assistant message, not a user message.
A user message can describe tone, style, or context. It is required for mimo-v2.5-tts-voicedesign.
You may omit the assistant message only when the voice-design model uses audio.optimize_text_preview: true, allowing the model to optimize or generate broadcast-ready text.

Audio Fields

Field	Type	Required	Description
`audio.format`	string	No	`wav`, `mp3`, `pcm`, or `pcm16`; non-streaming defaults to `wav`, while streaming should use `pcm16`
`audio.voice`	string	Model-dependent	A built-in voice name or Base64 voice sample for voice cloning
`audio.optimize_text_preview`	boolean	No	Voice-design model only; optimizes the target speech text; defaults to `false`
`stream`	boolean	No	Return audio chunks over SSE when `true`; defaults to `false`

Built-in voices include mimo_default, 冰糖, 茉莉, 苏打, 白桦, Mia, Chloe, Milo, and Dean.

Built-in Voice

curl -X POST "$MODELSELL_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $MODELSELL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mimo-v2.5-tts",
    "messages": [
      {
        "role": "user",
        "content": "Use a natural and warm tone."
      },
      {
        "role": "assistant",
        "content": "Welcome to the ModelSell speech synthesis service."
      }
    ],
    "audio": {
      "format": "wav",
      "voice": "Chloe"
    }
  }'

Voice Design

Voice design does not require an audio sample. Describe the voice in the user message and provide the target speech in the assistant message:

{
  "model": "mimo-v2.5-tts-voicedesign",
  "messages": [
    {
      "role": "user",
      "content": "A young male voice, clear and composed, with a moderate pace for a technology presentation."
    },
    {
      "role": "assistant",
      "content": "Explore an intelligent world where every conversation feels natural."
    }
  ],
  "audio": {
    "format": "wav"
  }
}

When "optimize_text_preview": true is set, the response also includes final_text_preview with the optimized target speech.

Voice Cloning

For voice cloning, pass Base64 from an MP3 or WAV sample in audio.voice:

{
  "model": "mimo-v2.5-tts-voiceclone",
  "messages": [
    {
      "role": "assistant",
      "content": "This sentence is generated with a cloned voice."
    }
  ],
  "audio": {
    "format": "wav",
    "voice": "BASE64_VOICE_SAMPLE"
  }
}

Read the Audio Response

For a non-streaming response, read the audio from choices[0].message.audio.data:

{
  "id": "chatcmpl_xxxxxxxxxxxxxxxx",
  "object": "chat.completion",
  "model": "mimo-v2.5-tts",
  "choices": [
    {
      "index": 0,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": "",
        "audio": {
          "id": "audio_xxxxxxxxxxxxxxxx",
          "data": "BASE64_AUDIO_DATA",
          "expires_at": null,
          "transcript": null
        }
      }
    }
  ]
}

Decode the Base64 into the requested format:

jq -r '.choices[0].message.audio.data' response.json | base64 --decode > output.wav

Streaming Audio

For streaming, set "stream": true and "audio": {"format": "pcm16"}. Each SSE chunk contains audio in choices[0].delta.audio.data. Base64-decode and concatenate the chunks in order to produce 24 kHz, PCM16LE, mono audio.

mimo-v2.5-tts supports low-latency streaming. Voice-design and voice-cloning models may return their streaming-formatted result only after inference finishes, so do not rely on them for real-time playback.

Model availability depends on the models enabled for your group on the current ModelSell site.