git:20260831.24816a9 to git:20260905.c1b14e1

---

name: audio-production

description: Produce exact-text dialogue speech clips, dialogue mixes, and final AV from approved Drama Dialogue and immutable source video. Use for provider-neutral voice resolution, speech freshness, duration probing, review, and deterministic AV assembly; do not use for authoring or revising Dialogue.

---

# Audio Production

Load the [Audio Layer content convention](../../docs/audio-layer-content-convention.md), the [Audio Projection contract](../../docs/audio-projection-contract.md), and the [scene-aware audio rules](references/scene-aware-audio.md). Treat the Audio Provider as a replaceable capability. Do not encode a vendor, transport, server, workflow, preset voice, or provider-specific pronunciation syntax in this Skill or in Drama content.

For video-conditioned final dubbing (Batch 7.3D), also read the [video-conditioned contract](../../docs/video-conditioned-audio-contract.md). This phase overrides the later Voice-design and AV-assembly steps: reuse frozen Video/Voice and accepted RealizedPerformanceSnapshot, keep DPD as dramatic-intent authority, and condition execution on actual visible facts. Preflight Video and Voice through the Drama Service owner with download/hash checks after storage migration; missing objects require reconciliation, never new identity. With unknown mouth activity, preserve natural delivery without guessed speech windows or filling the Video duration. Preserve a strict baseline separately; stop after durable dry dialogue and its review package, before lip sync, sound design, mixing or mux. Do not derive a new Character Voice Profile for this phase.

Gather the Scene with `scene.get_scene`, its Episode with `episode.get_episode`, its Script with `script.get_script`, its Work with `work.get_work`, and any requested Shot with `shot.get_shot`. Use `asset.search_assets` and `asset.get_asset` only when the Work/Scene chain does not already provide sufficient structured Character context; Character Assets do not require Media. Resolve each target only from `Scene.content.spokenContent[]`; preserve its exact `text`, `spokenContentId`, and `speakerKey`. A Provider adapter may derive a separately fingerprinted rendered-text representation using only officially documented and locally allowlisted controls, but it must preserve lexical content and never overwrite the canonical line. First produce an evidence-scoped, value-neutral Character Understanding with explicit unknowns; then derive a stable Character Voice Profile. When an approved `DPDSnapshot` is available, combine it with canonical SpokenContent, the stable Voice/Casting identity, the Voice baseline, and necessary Timing Context to create one typed `AudioPerformanceBrief`. The brief is the sole performance authority for the new path; do not also send Scene State or legacy Performance Intent. Include the hierarchy, Shot, Dialogue, character, listener, confidence, and evidence references in the structured request. Resolve applicable `Work.content.pronunciationGuidance[]`; stop on insufficient identity, DPD objective/relationship, Voice baseline, or unresolved pronunciation, and never repair gaps by changing Dialogue or DPD.

Compile the provider-neutral `SpeechGenerationRequest` with exact text in its typed field. Do not select a concrete provider, model, preset voice, reference identifier, transport, or endpoint. For the DPD Projection path, retain the platform Voice identity reference and the DPD/Projection fingerprints; the adapter reports every brief dimension as supported, approximated, or unsupported. It must not silently discard a direction or invent an unavailable parameter. The Role Dubbing Tool resolves `Work + speakerKey → voiceId → Voice`; the projected path requires that stable binding to match the brief. The legacy path may still reuse or materialize an approved Voice. For a new Voice, declare the stable use case (`CHARACTER_DIALOGUE` or `NARRATION`), reject every technically failing candidate before creative-fit comparison, and stop when the hash-addressed review package returns `VOICE_ARTISTIC_REVIEW_REQUIRED`. AI ranking is not approval. Only resume with an explicit user approval matching that package; do not redesign, import, materialize, bind, or synthesize before approval. A successful new binding identifies the durable Voice, never one Dialogue Media result.

Preflight only the required Role Dubbing and local probe capability. In contract-only, dry-run, or foundation work, stop before generation. In an explicitly authorized production run, call `production.generate_role_dubbing` once per stale spoken item with the structured request; never ask the Provider to write or improve the line. The Tool owns synthesis, output-side ASR, CER/missing/extra/repetition/proper-noun gates, durable Audio import, Voice identity, provider materialization, and Work binding. A failed intelligibility result cannot become formal Dialogue Media. A technically passing result remains `reviewStatus=PENDING` for human artistic Audio review while its technical review is PASS.

Inspect the durable identity with `voice.get_voice` when binding or provider-materialization evidence is required. Inspect the returned `ROLE_DUBBING_AUDIO` with `media.get_media` and resolve it with `media.resolve_media` when physical integrity or user review is required. Reconcile actual duration against the visual window without deleting, adding, or rewriting words. Prefer reviewed pace/pause or visual re-plan when needed. Deleting or invalidating a Dialogue Audio result never deletes its Voice; retiring a Voice never deletes historical Media.

Build one `av-assembly-v1` manifest. Reuse one speech clip across all relevant Shot slices; `spokenContentBindings` remains the visual coverage authority. An optional `SHOT_DIALOGUE_MIX` is derivative. Resolve durable inputs with `media.resolve_media`; never persist signed URLs. Mux to a new path, capture implementation/version/settings, probe both streams and duration, hash all inputs/output, and confirm the source Video hash is unchanged. If the capability is missing, return `AV_ASSEMBLY_CAPABILITY_MISSING` while preserving completed speech-foundation results.

Import a successful mux with `media.import_media` as a new `mediaType=VIDEO`, `purpose=FINAL_AV` Media identity with positive probed duration and the committed manifest/fingerprint. Resolve and hash-review it. A preview awaiting any MUST Audio review uses `final-av-attempt:<fingerprint>:<attempt-id>`; only a fully reviewed result uses `final-av:<fingerprint>`. Stop when the requested reviewed clips/mix/final AV are stable; do not begin another Scene or expand into sound design automatically.