AAA-tts-engine-catalog

AAA · TTS Engine Catalog

Five engines. Five trade-off axes. The right engine is a function of acoustic intent, latency budget, VRAM budget, and F13 sovereign gate.

DITEMPA BUKAN DIBERI.

What "Qualia" Means Here

F9 ANTI-HANTU floor — AI has no qualia. In this catalog, qualitative fidelity is defined mechanically as high-fidelity acoustic variance that mimics human biology:

• Pitch contour (F0 trajectory with micro-variations)
• Micro-pauses (50-300 ms silences inside phrases)
• Breath sounds (inhale, exhale, hesitation)
• Timbre shifts (emotional coloring)
• Tempo variation (WPM jitter, not constant rate)

P(Truth) of any qualitative claim: bounded by the F7 confidence cap (0.90) — listeners disagree.

The Five Engines

1. ChatTTS — Highest Qualia, Conversational

Use when: agent-to-human voice notes, conversational AI, expressive narration. The only engine that natively produces laughter and breath.

Avoid when: cloning i-ARIF (no good zero-shot path); multilingual mixing; deterministic output (varies by run).

2. F5-TTS / E2-TTS — Zero-Shot, Flow Matching

Use when: cloning a voice from a short sample; real-time-ish narration; low VRAM budget.

Avoid when: you need explicit emotional control tokens (no API for that — prosody is locked to the reference audio).

3. Fish Speech — Audio Language Model

Use when: stable multilingual TTS with code-switching in one utterance; long-form narration; semantic emotion matters more than prompt emotion.

Avoid when: VRAM constrained (< 8 GB); latency critical (< 200 ms).

4. MiniMax speech-2.8-hd — HD Commercial

Use when: HD narration for media artifact; i-ARIF voice (already enrolled via companion skill); high-stakes audio where quality wins.

Avoid when: budget constrained; local-only operation; F13 concerns about cloud-only identity storage.

See: /root/AAA/skills/AAA-voice-cloning-mimo-minimax/SKILL.md.

5. Edge / Mulberry — Free Fallback

Use when: free fallback when MiniMax / Qwen quota exhausted; voice bubble in chat UI; low-stakes narration.

Avoid when: identity continuity matters (no cloning); HD fidelity needed.

See: /root/.kimi-code/skills/tts-edge-fallback/SKILL.md.

6. MiMo Token Plan — Credit Economics & F13 Scope

The Xiaomi MiMo V2.5 family is the federation's primary subscription-backed multimodal surface (text + image + audio). Quoted from the official Token Plan pricing page — https://mimo.mi.com/docs/en-US/price/token-plan (probed 2026-08-19, F2 cite).

Catalog (6 active models)

Deprecated 2026-06-30: mimo-v2-pro, mimo-v2-omni, mimo-v2-tts (V2 family sunset).

Credit Deduction (per Token — language models)

ASR pricing

mimo-v2.5-asr is billed by audio duration: 30 M credits / hour (seconds-precision, rounded up to hours).

TTS pricing

All TTS series (mimo-v2.5-tts, ...-voiceclone, ...-voicedesign) are free for a limited time — do not deduct package credits. Use them aggressively.

Wire contract (verified live 2026-09-07): MiMo TTS rides POST /v1/chat/completions — assistant message = text to speak, user message = style instructions, audio{format, voice} field; response audio in choices[0].message.audio.data (base64 wav; streaming = 24kHz PCM16LE mono via delta.audio.data). NOT /v1/audio/speech. Gotchas: audio.voice is model-specific (omit entirely for voicedesign → HTTP 400 otherwise; voiceclone needs data-URI sample); builtin voices mimo_default/Mia/Chloe/Milo/Dean + 4 CN voices; singing via (唱歌) prefix on base tts only. Full contract: SOT mimo/mimo-v2.5-tts*.api_contract; deep reference: Hermes KVM4 skills/media/tts-edge-fallback/references/mimo-tts-api-quirks.md.

Plan tiers (monthly credits)

Annual tiers: ~12× monthly volume at 88% off (Lite $63.36/yr). First-purchase discount: 12% off (one-time per account, not on annual). Off-peak discount: 0.8× consumption during BJT 00:00–08:00 (UTC 16:00–24:00).

Endpoints

F13 SOVEREIGN Scope (license-binding)

Verbatim from the Token Plan page:

"The Token Plan package quota can only be used in programming tools (such as OpenClaw, OpenCode, etc.), and is prohibited from being used in the form of API calls for request behaviors in clearly non-Coding scenarios such as automated scripts and custom application backends."

Consequence for arifOS: All production Token Plan traffic MUST flow through a Forge Instrument harness (opencode, claude, codex, kimi, qwen, grok) or the OpenClaw gateway. Direct curl/SDK calls from custom backends or batch scripts are PROHIBITED under the license. Connectivity probes for testing are permitted (F2) but MUST be tagged [PROBE-NON-PRODUCTION] and not exercised in CI/cron loops.

Δ·Ω·Ψ meta-mesa (multimodal coherence)

Trade-off Matrix (canonical)

Routing Decision Tree (AAA-audio-emd-pipeline Phase 3)

Is the acoustic_intent a qualia token like [laugh] or [breath]?
├── YES → ChatTTS
└── NO  → Is there a sovereign voice_id to use?
         ├── YES → MiniMax (F13-gated identity)
         └── NO  → Is the deployment latency-critical realtime?
                  ├── YES (< 500 ms) → Edge / Mulberry
                  └── NO  → Is the sample short (3–10 s)?
                           ├── YES → F5-TTS
                           └── NO  → Multilingual mixing?
                                    ├── YES → Fish Speech
                                    └── NO  → F5-TTS (default)

EMD Integration (Phase 3 ENCODE)

Per AAA-audio-emd-pipeline:

audio = arifos.audio.tts_encode(
    decision_packet=packet,
    engine="chattts",          # or "f5-tts", "fish", "minimax", "edge"
    qualia_tokens=packet.acoustic_intent.breath_tokens,  # ChatTTS only
    voice_id="i-ARIF-2026-08-18",  # MiniMax only
    reference_audio="path/to/i-ARIF-15s.wav",  # F5-TTS, Fish
    output_format="mp3",
    sample_rate=24000,
)

Engine Selection by Use Case

ΔS Discipline

Per AAA-audio-emd-pipeline Phase 3:

• The encoder may not introduce qualia not authorized by acoustic_intent.
• [laugh] token in acoustic_intent MUST be matched by an engine that honors the token (ChatTTS). Routing to F5-TTS silently drops the token — return VOID and surface.

F13 SOVEREIGN Gates

When to Load This Skill

• Choosing a TTS engine for a new pipeline.
• Auditing engine selection in an existing pipeline.
• Switching engines after quota / latency issue.
• Wiring qualia tokens ([laugh], [breath]) into agent output.
• Deciding local vs cloud for VRAM / sovereignty reasons.

Integration Points

• Doctrine parent: /root/AAA/skills/AGI-audio-quantum-cognition/SKILL.md
• EMD pipeline (Phase 3): /root/AAA/skills/AAA-audio-emd-pipeline/SKILL.md
• MiniMax binding: /root/AAA/skills/AAA-voice-cloning-mimo-minimax/SKILL.md
• Qwen binding: /root/AAA/skills/AAA-voice-cloning-qwen-cloud/SKILL.md
• ASR binding: /root/AAA/skills/AAA-asr-glm-ingest/SKILL.md
• Edge fallback: /root/.kimi-code/skills/tts-edge-fallback/SKILL.md
• Audio analysis: /root/.agents/skills/audio-analysis/SKILL.md

Related Skills

• AGI-audio-quantum-cognition — Physics + floors
• AAA-audio-emd-pipeline — Phase 3 orchestration
• AGI-multimodal-bridge — Cross-modal evidence
• music-intelligence — Governed music generation (sibling domain)
• aaa-pdf-voice-protocol — Federation → human translation

Catalog forged 2026-08-18. F2 evidence: derived from open-source papers (ChatTTS, F5-TTS, E2-TTS, Fish Speech) + MiniMax technical spec + Edge TTS documentation. F7 confidence cap applies to all qualitative claims.