Chapter 14 — Sound after lock
Status: DRAFT 1 (2026-09-21), DESIGN v6 §4, §11. Sources: PRODUCTION.md:761–775; production/phase3/post-submission/SCORE-PLAN.md; production/phase3/score/mix_score.py, musicgen_cues.py; ledger rows 506, 524, 537, 540–541, 557, 566, 575–578, 593.
14.0 Objective
Given the locked cut, you can run a spotting pass first, then clean and unify dialogue, place laughs and breaths, lay ambience, and duck an AI score under dialogue, storing every audio file where the destination's verifier requires, with the tool named.
14.1 What you are deciding
What the film sounds like, in one pass of decisions before any file is made. In a generated film, sound "makes discontinuous pictures into one place" (PRODUCTION.md:761–763), and "one unbroken atmosphere laid under each location, running through the cuts … is the single highest-leverage sound decision in a generated film" (:767–770). The spotting pass (v6, on the director review) is you and the locked cut, once through with a pen: where music enters and leaves, which cuts the ambience must bridge, which lines need lifting, which breaths go. Then the layers in order: ambience, spot effects, dialogue, music last (:765, :775).
14.2 How we did it
Dialogue unification. Rock's first two lines had been generated with the platform's own voice; every later line was the chosen preset. A voice-change job converted the two, keeping timing ("Rock voice in 1A and 1C now the same", row 559), at 0.1 credit per second rounded up: 1 credit for 4.06 s (row 540), 2 for 10.08 s (row 557). Only the line window was mixed back; the ambience under the converted stretch turned to mumble, so the rest stayed native.
Breaths. The 6E line had been generated 15 dB under the others (row 503). Lifting the whole window by 14 dB (v1, row 506) lifted her breaths with it; the Director: "Crystal sigh too much". Five versions followed, each a measured change: v2 narrowed the lift to the line only (row 524); v3 ducked three breath windows found on a 50 ms RMS map (row 537); v4 replaced the bed with a looped clean stretch, "All Crystal breaths removed per Director" (row 541, "去掉 Crystal 的呼吸"); v5 replaced that bed, which "still carried breath texture", with a 150 Hz low-passed stretch (row 566). Every version has its window RMS numbers in the row.
Score. The platform had no music model (row 575). A local model was run on the GPU box (SCORE-PLAN.md:3), five cues on the scene boundaries of the locked cut (C1 0–26.1 s … C5 134.4–174.2 s, :9–13), one prompt per cue, two candidates per cue from the same prompt with different seeds (musicgen_cues.py:22). The Director picked B for all five ("A/B = same prompt, different seed", SCORE-PLAN.md:18), then the level: −20 %, then −10 % more on C1–C4 (0.72 linear), C5 untouched from the car scene to the end. The mix: each cue normalised to −26 LUFS, crossfaded to scene length, and side-chain compressed by the film's own audio so it dips under every line (mix_score.py:14–31). The five cue files were uploaded into the platform project (assets 266 → 271, row 593), as the rules required for any audio.
14.3 Case 14-A — four rebuilds for a breath
- What I saw: rows 506, 524, 537, 541, 566: five audio versions of one 12 s shot, four Director notes ("sigh too much", "breath still too loud", "remove Crystal breath entirely", "still breathing"), each version a measured window change with before/after RMS.
- Why I think so: the first fix was a window too wide (1.2–3.9 s for a 0.8 s line), and each later fix corrected the previous fix's side effect. No spotting pass had said "no breath in this shot" before v1; the rule arrived at v4.
- What I changed: the spotting pass as the first step, with a per-shot line for breaths and laughs; and the habit of measuring the window before and after every change (a 50 ms RMS map found the breath v3 needed).
- Would this happen on your tool? Every generated line comes with breath and room; every fix moves them. Decide on paper which you keep before the first filter runs.
14.4 Do it with your tools
| Need | We used | Any tool | If yours lacks it |
|---|---|---|---|
| AI score | a local music model on a private GPU | any hosted AI music tool, free tier or paid | required: some AI music source; the GPU is not on the required path (Appendix B) |
| Voice change keeping timing | the platform's voice-change job, 0.1/s | any voice conversion | re-record the line with the chosen voice and re-cut |
| Ducking | ffmpeg sidechaincompress keyed by the dialogue |
any editor's side-chain or manual volume automation | manual dips at every line, from the cue list |
| Window measurement | numpy RMS over the wav in 50 ms steps | any DAW meter | — |
14.5 Your turn
Time cap: two weekends. Spend: voice change and score rows of your sheet; ours were 3 credits and zero platform credits.
- Spotting pass on the locked cut, pen and paper: per scene, ambience source; per cut, bridge or not; per line, level note; per shot, breaths and laughs keep or remove; music in and out points; where the score must not be (ours: not under the packshot).
- Ambience first: one bed per location, running through every cut of that location.
- Dialogue: unify voices where a line was generated with the wrong one; lift the quiet ones on the line window only; remove what the spotting pass said; measure every window before and after.
- Laughs and breaths placed on visible laughs and removed elsewhere, per the pass.
- Score: one prompt per cue on scene boundaries, two seeds, pick by ear with the second reader, normalise, duck under dialogue, end before any required tail card. Ledger the tool.
- Store every audio file where the destination requires; name the tool in the row.
14.6 Gate
Each line needs valid evidence per DESIGN §5; outcomes are pass / fail / unverified, and unverified blocks.
- The spotting pass exists on paper, dated before the first post-lock sound file (the Chapter 7 line recordings predate it by design) (evidence: the file and the ledger order).
- One ambience bed per location, continuous across that location's cuts (evidence: the mix's track list; self-attested listening across each cut).
- Every dialogue change has a window measurement before and after (evidence: the rows).
- Score cues on scene boundaries, level set by ear, ducked under dialogue, with the tool named (evidence: the cue files, the mix command, the ledger).
- Every audio file stored where the verifier looks (evidence: the asset list).
- ★ The second reader listened to the full mix once, eyes closed, and named any moment the sound drew attention to itself (evidence: their note).
- Performance task: given our five 6E rows (506, 524, 537, 541, 566), write the spotting line for that shot as you would have written it before v1, then say which of the five versions it would still have needed and why (evidence: the line and the sentence).
Pass example from HALFWAY: the score, five cues, A/B by seed, level in two Director steps, ducked, uploaded. Fail example: 6E, five versions of one shot's sound because nobody had written "no breath" first.