Sound design: sound leads, picture follows, nothing stops abruptly

Viewers forgive a soft picture; they do not forgive bad sound. In a real production the client's main note on the final cut was "abrupt stops in many places", and an audit found the mechanisms: a title line longer than its fixed shot, about 2.5 s of dead air at most chapter ends, then the next sound cut off 0.3 s in. Every rule below exists to stop that class of problem, and the tools enforce most of them.

Sources, in order of preference

1. The client's own sound: the real sound inside their clips (python tools/archive/sound_events.py finds sustained tones like a conch or a bell, hits like drums and claps, laughter and chatter), and recordings they make for the film (hums, lullabies, ritual sounds, blessings; recording guides in the video-brief inbox README). A family's own conch blown on the morning of a ceremony beats any library conch, which may be from another tradition and sound wrong.
2. Licensed libraries, chosen by use: CC0 packs (e.g. Kenney) for anything; personal/educational-only archives (e.g. BBC Sound Effects RemArc) for private family films only, never client work. Log every file's source and licence in sources.md.
3. Music: public-domain melodies played fresh (a traditional tune or a pre-1929 composition, new performance), licensed library music (keep the licence PDF), commissioned music, or AI music whose terms allow the use (some carry watermarks such as SynthID; check commercial terms per tool). Never a commercial song without sync and master licences, even in a private film that may later be posted.
4. Synthesis (python tools/mix/synth.py): drones (tanpura bed), bells, low thumps, risers, room tone, pen scratches. Only for what has no real recording; filter it (a Karplus-Strong tanpura was hashy until a ~2.8 kHz low-pass) and check the spectrum and your ears.

Workflow

1. Spot the film with the storyboard page: per chapter, the music (or none), the real sounds, the effects, where the room should hear silence. Write it as comments in film/cues.py first.
2. Music-first edit. Pick one piece per chapter and keep it unbroken; cut picture to its phrases (python tools/mix/beats.py <music> --snap <t> gives beats, bars, phrase starts and accents). Land hits and bells on accents; tune bells to the key; edit music only on matching beats. A score chopped at every cut sounded broken in a look test; one piece with picture cut to it fixed it.
3. Write the cue sheet in film/cues.py → build(S). Place every cue relative to the timeline so it moves when narration changes:
   • S.T(chapter, shot, sec): a second inside a shot; S.VOAT(chapter, shot, k, sec) / S.VOEND(...): relative to the start / end of line k; S.dur(chapter[, shot]).
   • S.mus, S.amb, S.sfx, S.dia (clip sound, with clip EQ), S.voc (sung or hummed voice), S.loop(bus, path, at, dur, gain) for beds. Each has a default gain and a minimum fade-out (effects 0.15 s, dialogue 0.3 s, ambience 1 s, music 1.5 s) raised automatically; hard=True only for an intended hard stop.
   • Automatic: narration (polished, ducks everything), clip sound for shots with "sound": true, and a quiet room tone under the whole film so nothing drops to digital silence.
4. Bridge every chapter boundary: a ring-out (bell, reverb tail, held note) across the cut or a pre-lap of the next chapter's sound ~0.5 s early. Never silence-then-sound. Chapter stems are mixed with a window from -2 s to +14 s, so tails ring into the next chapter.
5. Mix: python tools/mix/film_mix.py (all) or ... film_mix.py e2 (one chapter, then the master is rebuilt from all stems; chapters whose start or length moved since their last mix are re-mixed too). Levels: narration on top; music bed -9 dB and ducked ~17 dB under narration; ambience -10, clip dialogue -14, effects -6 under narration. Two-pass loudness normalisation to the target in config mix (lufs, tp).
6. Check: python tools/qc/qc_audio.py flags DROP (audible to near-silent within 0.3 s, not at a line end), CUT TAIL (a drop right after a cut), DEAD AIR (> 1.5 s under -55 dBFS) and loudness > 1 LU off target; the mixer prints speech-over-bed per chapter (keep the lowest lines ≥ ~10 dB). Fix every flag or record why it is intended. Then listen on the delivery device, not only headphones.

Loudness by destination

Craft rules that held up

• Sound leads picture: the next shot's sound starts ~0.5 s before its cut.
• 1.5 s of quiet (room tone only) after an emotional line; music never swells under words.
• Real sound comes up full when the narrator is silent; let a laugh or a conch play out.
• Sing-alongs: word highlights aligned to measured onsets, not estimates (they were 0.7 s late once).
• Small speakers: nothing important below 120 Hz; presence around 2-4 kHz for narration; avoid harsh 6-8 kHz.
• A repeated sound motif (a heartbeat, an engine, a bell) at each chapter break ties the film together.

References

• references/cue-recipes.md: cue sheet recipes (openings, bridges, sing-along, montage, interval, ending)
• ${CLAUDE_PLUGIN_ROOT}/docs/compliance-and-licensing.md sections 1 and 3.4 (licences, music)
• ${CLAUDE_PLUGIN_ROOT}/docs/lessons-learned.md section E