Effects made in Python, free, and mixed by the compositor. They come out clean and a
little cartoony, which suits kinetic type. For realistic foley or a sound described in
words, generate it with elevenlabs-sound-effects (guide-add-sound-effects weighs the
two), or use the person's own files, placed at the same cues. If the person has brand
sounds or a sonic logo, use them in place of the synthesized ones.
1. Cues from the spec, not by ear
Have the build script write the cue list as it writes the spec, so a re-timed spec
re-cues the sound: a layer's start (entrance), end (exit), the key where a box
height reaches its minimum (flat), the contact time where x keys start moving (hit),
each with the x where its graphic sits. One sound per visible event; leave space
between them, and vary the sound by kind of graphic (one sound repeated 8+ times in 20 s
grates).
2. The palette
references/sfx.md holds 24 recipes and the mixer. Pick by event:
| Event | Sounds |
|---|---|
| A title or badge pops in; blinks out | pop; down_pop |
| A label, a cursor, a counter step | click, tick; type_key repeated at the typing's rate |
| A title card, an end card | blip (8-bit), ding on a check or "done" |
| A highlight or reveal | sparkle; riser into it |
| A word slams down, a stamp lands, the hero word | impact, stamp, bass_drop |
| A hand hits a title; it's squashed; pressed flat | thwack; squash; splat |
| Something drops on a head, bounces | bonk, boing |
| Something flies past or off | whoosh, panned with to_x from the hit side to the exit side |
| A stop, a glitch, a slice | record_scratch, glitch, snip |
| Applause, confetti, a title crushed | clap, cannon, crumple |
The six recipes the first job tuned by ear: a pop is a sine sweep from about 780 Hz up to 1600 Hz over 90 ms with an exponential decay (~38/s) and a 2–6 kHz noise tick; the squash a sawtooth from 260 Hz down to 70 Hz with an 18 Hz wobble, low-passed at 1.4 kHz, with a gated band-passed creak, getting louder as it flattens; the splat a 90 Hz falling thump (~14/s) with low-passed noise; the down-pop a sine from 900 to 300 Hz over 120 ms, quieter; the thwack a 140 Hz falling thump with a 0.8–5 kHz noise burst over 250 ms; the whoosh noise through a band-pass whose centre sweeps 400 Hz to 4 kHz over ~0.45 s.
3. Build the track
Write references/sfx.md's script into the job's scripts/ as sfx.py, as is; the build
script writes work/sfx_cues.json (the format is in the reference) and then:
exec python scripts/sfx.py build work/sfx_cues.json audio/sfx.wav
Every sound's transient lands 10 ms before its cue, panned by its graphic's x (equal-power, ±0.8 at the frame edges); the sum is soft-limited where sounds stack and normalized to a −1 dBFS peak, 48 kHz stereo, the comp's exact length. The noise is seeded, so a rebuild sounds the same.
4. Mix it
In the spec, the effects under the voice:
"audio": [{"src": "../audio/voice.wav", "start": 0, "trim": 0, "volume": 1}, {"src": "../audio/sfx.wav", "start": 0, "trim": 0, "volume": 0.6}].
With no voice (effects in place of the original sound), the effects alone at 1. To keep
the original sound under the effects, list the footage as its own audio entry at a lower
volume. Changing audio needs a whole render, not --into.
5. Check
- After the render, run
volumedetecton a stretch with no events: about -91 dB means the original audio is really gone. - You can't hear: put the render to the person with
review_approve, each sound and its moment inbrief, and ask them to listen. Adjust gain or swap a recipe per their notes, then rebuild from the script.


