ariffazil MiMo Audio — federated organ (AAA canonical, 2026-09-07)
Status: LIVE, E2E-verified. TTS = 0 credits while free window lasts.
Env (every FI home + Hermes): MIMO_TOKEN_PLAN_API_KEY + MIMO_BASE_URL (https://token-plan-sgp.xiaomimimo.com/v1); alias MIMO_API_KEY.
Contracts SOT: /root/.config/federation-models.json → mimo/mimo-v2.5-* (api_contract) + signatures fed-realtime-voice (WIRED_DIRECT_LANE) / fed-audio-understanding.
⚠️ TRANSPORT WARNING (E2E-tested 2026-09-07)
- FED front door
:4000lists all 6 mimo models (discovery GREEN) — but litellm's translator CANNOT carry audio bodies: TTS response crashesconvert_dict_to_response.py(message.audiounknown); ASRinput_audiocontent part gets mangled → provider "Param Incorrect". - Audio bodies MUST ride the DIRECT lane (agent →
$MIMO_BASE_URL/chat/completions). ToS-clean agent-tool usage, E2E-proven below. Text-lane MiMo traffic may use :4000 normally. - Filed for FED owner (FI-008): litellm audio-body support + classifier audio lexicon (v1 has no audio class).
The three TTS models — ALL ride POST {BASE}/chat/completions (NOT /audio/speech)
| model | voice source | singing | clone | design |
|---|---|---|---|---|
| mimo-v2.5-tts | builtin | yes ((唱歌)/(sing) prefix, CN lyrics best) | no | no |
| mimo-v2.5-tts-voicedesign | text description (user msg) | no | no | yes |
| mimo-v2.5-tts-voiceclone | audio sample data-URI <=10MB mp3/wav | no | yes | no |
Wire: assistant message = text to speak; user message = style instructions (not spoken); audio: {format: "wav"|"pcm16"(stream), voice: "<id>"}; response choices[0].message.audio.data = base64 wav; stream = delta.audio.data 24kHz PCM16LE mono.
Voices: EN Mia / Chloe / Milo / Dean — CN 冰糖 / 茉莉 / 苏打 / 白桦 (+ mimo_default).
Style control: natural language (user msg) + inline audio tags [pause] [sighs] [sternly] [trembling] (bilingual, mixable) + screenplay layers Instruct/Text/Character/Scene/Direction.
Gotchas: audio.voice is model-specific — voicedesign MUST omit it (HTTP 400); clone needs data-URI; TTS output is random — regenerate a few times to pick the best take.
ASR — mimo-v2.5-asr (chat/completions, NOT /audio/transcriptions)
Content part: {"type":"input_audio","input_audio":{"data":"data:audio/wav;base64,..."}}; wav/mp3 ONLY; b64 <= 10MB; asr_options.language in auto|zh|en — no ms; BM/dialect rides auto (Penang-Besi UNTESTED → prefer dashscope/GLM for BM fidelity). 30M credits/audio-hour ($0.074/h).
Governance
- voiceclone/voicedesign new voice_id = F13 SOVEREIGN gate; reuse = F11 audit (see AAA-voice-cloning-mimo-minimax).
- i-ARIF voice_id stays MiniMax speech-2.8-hd — do NOT move.
- ToS: token-plan key for agent/harness use only — never cron/backend.
Fallback ladder
mimo-audio (0 credits, free window) → edge-tts (free floor, ms-MY voices) → QwenCloud token-plan TTS.
E2E receipts (2026-09-07)
- Direct TTS (Hermes KVM4, Mia): 307,244B wav / 24kHz mono / 6.4s / 223 tokens / 0 credits
- Direct TTS (FI-003 KVM8): 133,180 b64 chars / 173 tokens / 0 credits
- Vision understanding: 128x128 red PNG → "Red." (img_tokens=16, cached=192)
- litellm :4000 TTS/ASR bodies: RED (translator) — warning above
Quick recipe (direct lane)
curl -sS "$MIMO_BASE_URL/chat/completions" \
-H "Authorization: Bearer $MIMO_TOKEN_PLAN_API_KEY" \
-H 'Content-Type: application/json' -d '{
"model":"mimo-v2.5-tts",
"messages":[{"role":"user","content":"Warm, friendly Malaysian English."},{"role":"assistant","content":"Good morning, the federation is healthy."}],
"audio":{"format":"wav","voice":"Mia"}}' \
| jq -r '.choices[0].message.audio.data' | base64 -d > out.wav
Federation map
Canonical: /root/AAA/skills/mimo-audio (this file) · Hermes original: KVM4 ~/.hermes/skills/media/mimo-audio · visible to: qwen/claude/codex/grok (whole-dir AAA links) + kimi-code/opencode-config/gemini/.agents (per-skill links) — 9 homes total · official vendor reference: references/official/ (MIT, XiaomiMiMo/MiMo-Skills).