openclaw - 💡(How to fix) Fix fix(tts/xiaomi): support per-call voice and model overrides [1 comments, 2 participants]

Summary

The built-in Xiaomi MiMo TTS provider can generate speech successfully, but it does not currently support per-call voice / model overrides through the standard OpenClaw TTS CLI/Gateway path.

This makes it hard to use MiMo TTS for interactive language-learning / pronunciation practice, where users often need to switch voices such as Dean, Milo, Mia, etc., or test newer MiMo V2.5 TTS model variants.

Environment

• OpenClaw: 2026.5.4 (325df3e)
• Provider: built-in xiaomi speech provider
• TTS endpoint: Xiaomi MiMo OpenAI-compatible chat-completions TTS API
• Active provider: xiaomi

Configured provider list reports Xiaomi correctly:

{
  "active": "xiaomi",
  "providers": [
    {
      "id": "xiaomi",
      "name": "Xiaomi MiMo",
      "configured": true,
      "selected": true,
      "models": ["mimo-v2.5-tts", "mimo-v2-tts"],
      "voices": ["mimo_default", "default_zh", "default_en", "Mia", "Chloe", "Milo", "Dean"]
    }
  ]
}

Reproduction

Run a one-off TTS conversion with a Xiaomi voice override:

openclaw capability tts convert \
  --local \
  --channel telegram \
  --voice Dean \
  --text "This is Dean. American English practice." \
  --output /tmp/mimo-dean.mp3 \
  --json

Actual result:

Error: TTS provider "xiaomi" does not support model or voice overrides.

The same happens with --model xiaomi/mimo-v2.5-tts and with Gateway transport.

Expected behavior

The Xiaomi provider should support standard per-call overrides:

• --voice Dean
• --voice Milo
• --model xiaomi/mimo-v2.5-tts
• eventually newer model variants such as:
  • MiMo-V2.5-TTS
  • MiMo-V2.5-TTS-VoiceDesign
  • MiMo-V2.5-TTS-VoiceClone

The provider source already appears to parse directive-style override keys internally (voice, voiceId, model, xiaomi_model, etc.), and the Xiaomi API itself accepts the audio.voice field. Direct API calls using audio: { format: "mp3", voice: "Dean" } work.

Use case

Interactive English practice over Telegram:

1. Assistant reads a sentence using Xiaomi MiMo TTS.
2. User repeats it by voice message.
3. Assistant provides pronunciation feedback.
4. User may want a male/female voice, or specifically standard American English style guidance.

For this workflow, per-call voice switching is important. Editing global config for each voice is too heavy and not suitable for chat-based practice.

Suggested fix

Please add Xiaomi support for explicit TTS overrides in the standard OpenClaw TTS path, for example by implementing / exposing the provider override resolver so CLI/Gateway calls can pass through:

• voice / voiceId → Xiaomi audio.voice
• model / modelId → Xiaomi request model
• optional style → user style message, if supported by the current runtime path

Also consider updating the built-in Xiaomi TTS model catalog and docs to clarify:

• supported voice ids
• whether voices imply US/UK accent or require style prompting
• support status for VoiceDesign and VoiceClone

Related

• #72927 appears related to MiMo V2.5 TTS VoiceDesign support.
• This issue is specifically about per-call voice / model overrides for the existing Xiaomi TTS provider.