AI Audio Sound Design

AI-generated clips have three tells that scream "fake" before a viewer can articulate why: dead silence between words (real rooms are never silent), dry studio voices in visibly reverberant spaces (a bathroom that doesn't sound like a bathroom), and a tonal watermark whine at 13–16 kHz that every Arcads-style generator leaves in the track. This skill fixes all three per scene, plus broadcast-style censor bleeps, without touching a single video frame.

Proven on a real 10-scene AI-actor ad (5 approved client revision rounds). The engine is scripts/build_mix.py; per-location recipes and level targets are in references/acoustics.md — read it before filling in the scene table.

What ships

• out/<name>-vN.mp4 — video stream copied bit-exact (-c:v copy), audio rebuilt (AAC 256k).
• Stems (-dialog, -amb, -bleeps wavs) — every level decision gets measured, not guessed.
• Review canvas per the video-review-canvas skill; new version = new filename, same slug.
• Deliver the finished file next to the source (same folder or cloud location it came from).