Soundtrack

Write the sound for this film

A film's sound is written for that film: its own instruments, rhythm and harmony in Web Audio, from nothing, in window.SCORE = ac => { … } (oscillators, noise, filters, envelopes, delays: whatever the piece calls for). Drums, a groove or a music-box bed are written the same way, for this film: no two films should sound alike because they share a template. templates/groove.js and templates/score.js are two ready-made scores, used only when the user asks for them; each file documents its options.

render.mjs (from the javascript-animation skill) renders window.SCORE offline with OfflineAudioContext, writes <out>.wav and muxes it into the MP4; the page's live preview plays the same score when clicked.

Don't ask the user to choose a music source up front. At delivery, mention in one line that AI-generated music or their own track also work.

How the music follows the picture

• Tempo map. Each scene gets a whole number of bars (the tempo may flex per scene), so every cut lands on a downbeat.
• Mark cuts with harmony, not volume. Each scene walks home to the dominant and the cut arrives on the tonic. A loud accent on a cut (bass + bright chime at once) reads as a jump scare: users flagged a +7 to +10 dB jump as startling; +3 to +5 dB with a chord change reads as a clean edit.
• Don't thin the opening bar if it's the only bar before the first cut; a sparse intro followed by the full texture is itself a startle.
• No extra accent on the cut's downbeat. Even the normal downbeat accent (louder bass + first note) tipped short pieces over +6 dB, because the bar before is decaying. The chord change is enough.
• Hard cuts inside a scene count. Put them in window.CUES too, so they get the same downbeat.
• Give the eye something to lock onto. Legato music with no percussion has no accents for cuts to land on. With a steady groove, accents are part of the rhythm and don't startle (the check compares against the last few seconds, not an isolated quiet moment).
• Bring layers in one or two per bar. Kick and bass entering together after a quiet intro measured +11 dB (startle); staggered by a bar, +3.5 dB.
• Use a break. Cutting everything to silence for a bar before the payoff is the strongest beat in a short. sync-check accepts a cut into silence as on the beat.

Other sources (only when asked)

• AI-generated music (e.g. Lyria via OpenRouter): see references/ai-music.md.
• The user's own track: mux it with ffmpeg.

With music you don't control, reverse the direction: the picture follows the music. Render, mux, run the check, then move the scene boundaries in the page's SCENES table by the reported offsets and re-render. This is cheap because each scene function receives local time (0 at its own start), as the starter is written: moving a boundary never touches scene code.

Always: check it

node scripts/sync-check.mjs piece.mp4 --cues-file piece.cues.json   # check the declared CUES (preferred)
node scripts/sync-check.mjs piece.mp4                       # fallback: detect cuts in the picture (fast camera moves can read as cuts)

For each event it prints the offset to the nearest strong audio onset and the loudness jump into it. It fails on OFF-BEAT (more than ±100 ms, ~3 frames) or STARTLE (more than +6 dB versus the second before). The opening second and the final fade are skipped (no "before" to compare against; a fade-out is not a cut). "Strong" means prominent among the onsets within ±2 s, so accents in quiet passages still count; a cut into silence counts as landing; music returning after a break is compared with the typical level of the last 4 s. A video with no audio stream is reported as such. Put in window.CUES every cut plus any in-scene hit you want checked (a hard cut inside a shot, a gag's impact). Fix and re-run until it passes, or until every remaining STARTLE is a deliberate stinger you can name (a comedy impact is meant to jump out; a lullaby cut is not).

Limits

Synthesized in code: music box, piano-like plucks, pads, bells, soft percussion are fine. Realistic voices, orchestras and genre production are not; use AI music or a real track for those.