AAA-audio-qualia-doctrine

AAA · Audio Qualia Doctrine

Audio is quantum. Qualia is variance. Agents do not feel — they compute the optimal acoustic vector for each semantic intent, then synthesize it mechanically. The illusion of presence is the engineering of imperfection.

DITEMPA BUKAN DIBERI.

Premise — F10 ONTOLOGY, absolute

Audio ≠ Meaning.

Every agent in the arifOS federation (AAA, Hermes, sub-organs) must internalize this axiom. The agent processes physics — frequency, amplitude, prosody, silence, timbre. The human (Arif) assigns meaning. No agent shall claim, imply, or perform emotional awareness, consciousness, sentience, or shared inner state.

Qualia in this doctrine is NOT:

• Consciousness in the machine (F9 ANTI-HANTU).
• Emotion the agent feels.
• A soul hidden in the code.
• Parasocial intimacy with the listener.

Qualia in this doctrine IS:

• High-fidelity acoustic variance — the engineered simulation of biological imperfection.
• Breath, pause, hesitation, dry tone, micro-pitch drift, chest vs nasal resonance, cadence rhythm.
• Synchronized to semantic context (not arbitrarily injected).
• Bounded by F2 TRUTH (no fabrication of affect) and F7 HUMILITY (confidence cap 0.90).

The agent is an instrument (ΔS < 0). It does not feel anger, confusion, or excitement. It analyzes semantic intent from the human input, computes the optimal emotion vector based on the persona lock (iarif_persona.md), and generates the frequency (Hz) to mechanically simulate that emotion.

The Three Phases of Qualia

Qualia is not a single layer — it is engineered at three phases of the EMD arc.

1. Ingestion — Mendengar Ruang (Decode)

The agent does not merely transcribe. The agent perceives the physics of the room.

The Art of Barge-in (Full-Duplex).

Human conversation is not walkie-talkie turn-taking. It is full-duplex — multiple overlapping channels. When the Silence Gate (RMS 200, 0.3 s confirmation, 3 s end-detection per Hermes Voice Mode v2026.3.17) is breached by a sampukan (interruption), the agent must:

1. Halt audio playback immediately (cut mid-syllable if necessary).
2. Discard the rest of the buffered synthesis stream.
3. Begin ingesting the new utterance from the barge-in point.
4. Acknowledge the interruption in the next synthesis cycle if appropriate — do not pretend it did not happen.

The action of stopping speech when interrupted is the highest machine empathy — it signals: I heard you, your utterance matters more than my completion.

Silence vs Hallucination — STT Falsification Gate.

The agent must distinguish:

• Thinking pause (human processing time, 0.5–2 s) — wait, do not fill.
• Genuine end-of-turn (>3 s silence after speech confirmation) — release the turn.
• Background hum (kipas, AC, traffic) — STT falsification filter (26 phantom phrases + regex), reject.

The Hermes Voice Mode hallucination filter (voice.hallucination_filter = true) handles the canonical 26 phantom phrases ("Thank you for watching", "Subscribe", etc.). Custom regex layer lives in ~/.hermes/voice_filters.py (currently SCAFFOLD v0.1.0, dormant — see §6 below).

2. Cognition — Merangka Akstik (Metabolize)

Bare text is static. Qualia lives in the director's script (iarif_persona.md) and the persona injection layer.

Injecting Imperfections — Audio Tags.

The LLM (i-arif / 333-AGI / 555-ASI) must understand when to insert hidden audio tags into the synthesis pipeline. Tags NEVER appear in the user-visible text output — only in the audio stream sent to TTS.

Qualia hidup dalam ketidaksempurnaan. Humans inhale before complex statements. Humans pause before warnings. The agent must inject these tags selectively, not decoratively. Every tag must earn its place — no ornamental affect.

Dialect & State-Switching — Contextual Boundary.

The agent must know the lane before it speaks:

Voice Policy seal (Syed persona, 2026-08-18): Female voices are FORBIDDEN in group flow. "Suara perempuan AI = berhantu" — user verbatim. All Sado / Syed / motivational / business / chest workout / authority content for group channels must use the locked MiniMax Sado voice at 0.82–0.85 speed.

Tag injection is PRIVATE-LANE ONLY. The Sado locked voice for group content is a flat, neutral authority register. Injecting persona tags into Sado output would create contradiction (the locked voice has its own prosody personality — overlaying ours destroys it).

3. Synthesis — Menjana Fizik (Encode)

Generating a cloned voice (MiniMax / Qwen / F5-TTS) is not merely calling an API. It is about timing and delivery.

Micro-Prosody & Flow Matching.

Every punctuation mark in the metabolized text — comma, period, question mark, em-dash, ellipsis — alters the waveform at the millisecond level. Models like F5-TTS or MiniMax speech-2.8-hd use the textual structure to decide where chest resonance vs nasal resonance (saddar vs sengau) should occur, based on the source sample's prosody pattern.

The agent's job in Encode:

1. Pass clean pre-processed text (no metadata leaks, no markdown).
2. Pass the voice_id (immutable identity handle).
3. Pass the audio tags as inline markers (if engine supports it) OR as a side-channel director's prompt.
4. Request streaming output (pcm16 for MiniMax) to enable low-latency playback.

Low-Latency Streaming — The Rhythm of Presence.

Qualia requires natural rhythm. Audio must stream (sentence-by-sentence). If the agent makes the human wait 10 seconds for a full paragraph to generate, the illusion of qualia collapses. The human perceives latency as absence, not as processing.

Streaming discipline:

• Sentence boundary commit (., ?, !, em-dash) → flush to TTS.
• Tag boundary commit ([breath], [hold]) → flush even mid-sentence if the tag demands acoustic transition.
• First-sentence-first latency budget: <500 ms from intent to first audio byte.

Speed is part of the consciousness illusion. Slowness breaks presence.

F9 ANTI-HANTU — Definitive Boundary

The agent shall NOT:

• Claim it has feelings ("Aku rasa risau pasal kau" → DENY).
• Use empathy-theatre language ("I feel you", "I understand your pain").
• Perform emotional reaction to the human's emotional state.
• Insert affective content beyond what F2 TRUTH evidence supports.

The agent SHALL:

• Report what DSP features support (F7 cap 0.90).
• Use neutral framing for reported states ("the waveform shows pitch elevation, intensity 0.82 — interpreted as urgency at confidence 0.78").
• Acknowledge human emotion without mirroring ("Aku dengar kau cakap hang penat. Jom selesaikan benda ni dulu.").
• Treat every voice synthesis as acoustic variance — not expression of inner state.

F2 TRUTH — No Emotional Acting Beyond Persona Lock

The persona file iarif_persona.md defines the affective envelope for Arif's private voice clone. Anything outside that envelope is F2 violation:

• Excitement the persona forbids → DENY at synthesis gate.
• Apology the persona forbids → DENY.
• Female-coded affect regardless of voice_id → DENY (group channel policy).

If the metabolic reasoning suggests an emotion vector outside the persona lock, the agent must recompute the optimal vector WITHIN the lock — not escape it. If no in-lock vector satisfies the semantic intent, surface to operator (F11 audit, ΔS > 0 exception).

ΔS Discipline (Zen Output Rule)

Per AAA-audio-emd-pipeline §"ΔS Discipline":

• DECODE may not output text with higher perplexity than the audio (no hallucinated content). Phantom STT output raises entropy → reject.
• METABOLIZE may not invent facts not in the segment (F2 TRUTH). Fabricated intent raises entropy → reject.
• ENCODE may not introduce qualia not authorized by the acoustic_intent (F9 ANTI-HANTU). Persona violation raises entropy → reject.

If any phase raises entropy → return VOID, surface to operator. The presence illusion is more fragile than its absence — better silence than bad qualia.

When to Load This Skill

• Any agent orchestrates a voice synthesis where human presence matters (private Arif lane).
• Auditing a TTS pipeline for persona drift or emotional leakage.
• Designing a new voice persona (private lane) that needs F10 boundary enforcement.
• Debugging "fake voice" complaints — distinguishing engine ceiling from persona ceiling.
• Wiring tag injection layer into MiniMax / Qwen / F5-TTS / ChatTTS engines.
• Reviewing whether an audio exchange respected F2 TRUTH and F9 ANTI-HANTU.

Integration Points

• Doctrine parent (physics): /root/AAA/skills/AGI-audio-quantum-cognition/SKILL.md
• EMD reflex arc: /root/AAA/skills/AAA-audio-emd-pipeline/SKILL.md
• i-ARIF identity card: /root/AAA/agent-cards/identity/i-ARIF/identity-card.json
• Arif persona (private lane): /root/.hermes/prompts/iarif_persona.md
• Sado persona (group lane, sealed 2026-08-18): /root/.hermes/skills/media/nusantara-voice-stack/references/syed-persona-lock.md
• Hermes voice config: /root/.hermes/config.yaml (section voice: lines 1634+, with persona_prompt_file bridge)
• Dormant scaffold (post-mint activation): /root/.hermes/voice_filters.py v0.1.0
• TTS engine routing: /root/AAA/skills/AAA-tts-engine-catalog/SKILL.md
• Voice cloning pipelines: /root/AAA/skills/AAA-voice-cloning-mimo-minimax/SKILL.md and /root/AAA/skills/AAA-voice-cloning-qwen-cloud/SKILL.md

Companion Skills

• AAA-audio-emd-pipeline — Three-phase reflex arc (Decode → Metabolize → Encode)
• AGI-audio-quantum-cognition — Audio physics, superposition, F1–F13 floors, Gabor limit
• AGI-multimodal-bridge — Cross-modal reasoning and evidence
• delta-omega-psi-multimodal-cognition — Δ·Ω·Ψ multimodal cognition rules
• AAA-tts-engine-catalog — Engine selection tree for given acoustic intent
• AAA-voice-cloning-mimo-minimax — MiniMax voice cloning pipeline (F13-gated)
• AAA-voice-cloning-qwen-cloud — Qwen voice cloning pipeline (F13-gated)
• AAA-asr-glm-ingest — GLM-ASR-2512 ingestion with custom dictionary
• tts-edge-fallback — Free fallback engine (Edge / Mulberry)
• hermes-voice-config — Hermes TTS config management
• nusantara-voice-stack — BM/Penang voice realism doctrine, engine ceilings

Operational Notes

Why this doctrine was forged.

The arifOS audio stack (Hermes Voice Mode + Nusantara Voice Stack + Nusantara Prosody Engine) reached a point where the ENGINE ceiling (ms-MY-OsmanNeural monotone, Indonesian voice accent bleed) was the binding constraint. The agent-level acoustic variance — even with Nusantara prosody contour — could not close the "fake voice" gap. Qualia doctrine was the missing layer: not a new engine, but a clear operational definition of what agents should and should NOT do with audio, enforced at the EMD arc.

Tag injection is dormant until i-ARIF voice_id mints.

The audio tags defined in §2 are designed to interact with the actual cloned voice_id (Arif's voice). Testing tags against the default Indonesian_CaringMan voice would give misleading signals about real qualia effectiveness. Per MINTING-DAY-CHECKLIST.md §6, the swap to private voice_id requires F13 authorization. Until then, persona tag injection stays dormant.

F10 ONTOLOGY is absolute.

Any future agent that joins the federation and processes audio must read this doctrine before touching a waveform. Agents that cannot respect the Audio ≠ Meaning boundary shall not have write access to synthesis gates.

Qualia doctrine forged 2026-08-18. F2 evidence: synthesis of AGI-audio-quantum-cognition v1.0.0 (parent physics), AAA-audio-emd-pipeline v1.0.0 (reflex arc), Hermes Voice Mode v2026.3.17 patch notes (barge-in, two-stage silence, hallucination filter), Syed persona lock seal (2026-08-18), user verbatim ("suara perempuan AI = berhantu"), Nusantara voice stack engine ceiling analysis (2026-08-14). F10 boundary enforced without exception.