produce-generative-audio

Produce Generative Audio

When to use

• Generating or evaluating music, sound effects, ambience, or voice candidates.
• Building a deterministic audio batch and listen-through review.
• Preparing accepted audio for a conventional runtime handoff.

Workflow

For SFX and ambience, route to build-audio-assets and audio-pipeline.sh; retain this entry point for modality dispatch and existing compatible media workflows. AUDIO has its own brief, immutable lifecycle, technical/listening/loop review and separate target/promotion boundaries, independent from MUSIC, VOICE and visual selection. Never fall back to a direct or different-provider workflow when a selected domain operation fails.

For instrumental or supplied-lyric music, route to build-music-assets and music-pipeline.sh. Its exact lyric-rights identities, disabled planning/normalization, full-track/line review and separate current target/distribution gates extend the shared lifecycle without importing AUDIO. Preserve old failed song attempts; an instrumental result cannot satisfy requested singing.

For preset speech or consented local cloning, route to build-voice-assets and voice-pipeline.sh. Native consent and host local-use evidence bind the actual task/purpose/reference; public output disclosure and target distribution are separate. Keep recordings, transcripts, conditioning, generated speech and identifying hashes private. No cloud cloning/upload, automatic transcript or runtime promotion is implied.

1. Confirm the selected backend and audio capability in MEDIA.md. For LOCAL, use use-bubo-local-generation through bubo-media.sh; read Bubo's media API for exact formats/settings and intrinsic decoder behavior. Setup, zero-sample diagnostics and generation are separate operations.
2. Define purpose, duration/bounds, structure, loop intent, variants, sample rate, channels, loudness, peaks, file size and runtime target. Direct Bubo LOCAL distinguishes SFX, instrumental/supplied-lyric music, preset speech and consented cloning. Voice uses explicit autoregressive token/time/byte ceilings rather than a fixed requested duration or diffusion controls; unsupported settings must fail.
3. Record source rights and obtain explicit identity or voice consent where applicable. Cloning requires exact local reference audio/transcript identities and reference-bound rights/consent; reject remote reference URLs. Keep transcripts/conditioning private and separate local-use consent from public audio disclosure.
4. Generate a deterministic batch where supported and record explicit unavailable markers where a provider withholds parameters or seeds. Preserve the approval's shared media-state path and existing candidate/revision allowance across tasks and workspace changes.
5. Stage candidates and create a listen-through sheet with waveform, duration, loudness, clipping, silence, loop, artifact and semantic checks. Keep native amplitudes and disclose clipping/silence rather than hiding them with normalization.
6. Reject unsuitable candidates explicitly. Promote only accepted files after human review and runtime validation.

Output

• An audio brief, batch plan, and runtime budget.
• Staged candidates with provenance and generation manifests.
• A listen-through review with acceptance or rejection reasons.

Guardrails

• Never clone or imply a person's voice without documented authorization and consent.
• Do not assume generated music, samples, or model outputs are redistributable.
• Do not normalize, trim, splice, loop, or repair a rejected artifact destructively without approval.
• Do not publish or promote audio automatically.
• Listen to the complete candidate for content and declared preset/reference match. A settings manifest, waveform measurement or fixture is not a listening verdict; generated music never implies a reviewed loop.