A library sound placed on a frame is still a guess: it has a pitch, a length, and a position that were decided somewhere else. When the sound feels wrong against the animation, it is almost always one of those three disagreeing with what the eye sees — a small chip landing with a low thud, a card on the right edge sounding from the centre, a whoosh that peaks on a frame where nothing is moving fast. The fix is not a better library. It is to derive the sound from the motion and let the picture decide.

The mapping

The motion says	The sound does	Rule of thumb
Size (√ of the element's area)	pitch and decay	60 px → ~5 kHz and ~165 ms (a `tick`, 50 ms); a 460 px headline → ~1.1 kHz and 600 ms. `f0 = 5200 · (60 / size)^0.75`, `t60 = 0.1 + size / 460 · 0.5`. In the bed-and-clicks register (B), clicks stay in 2.8–5.2 kHz whatever the size: a 20 ms sound has no room for a low fundamental, so a big thing gets a thud on the bed's root instead. Product-camera foley (E) sits lower, at 1.45–2.5 kHz (Linear Agent, Raycast); set `pitch` per cue, since the derived click band floors at 2.8 kHz
x on the canvas	stereo pan	constant-power, `pan = (x / W − 0.5) · 1.4`, clamped to ±0.7 — never hard left or right; thuds pan at 40 %
y on the canvas	brightness of the attack	higher on screen = brighter contact noise; the ear reads height as brightness, not as pitch
Direction of travel	contour	leaving upward → the breath sweeps up; sliding in from the left → the tick's pan starts left of where it lands
Tween duration and ease	length of the breath, and where it peaks	a whoosh starts with the tween and peaks on the fastest frame: the cut for an exit that accelerates into it, mid-travel for an ease-in-out cross, near the start for a decelerating entrance, whose settle gets its own `land` or `thud`. Loudness of a moving object follows speed, so a peak on the settle sounds backwards (designed-sfx-anatomy)
A stepped reveal	one click per step	a glyph flipbook at seven frames a step, words at 110 ms, table cells at 210 ms — the cadence is the sound
Level	six dB per doubling of size	inside a narrow window so the biggest thing is loudest but nothing shouts. In register B a click's median peak at the hit runs from about level with the sub (Refreshed. +1.7 dB) to 7 dB under it (GPT-5 −7.3 dB); across the film, at p95, the high band sits ~19 dB under the sub band. Where foley follows consequence (E), commits outrank typing — launch-video-sound
A reveal or an act change	the bed's or track's structure	the drop or the low end returns on the reveal frame, and something leaves 0.25–0.9 s before it (the sub in Contra Indy, the whole mix in Taste Labs); a key change can mark an act (Linear Agent) — launch-video-music, launch-video-sound

Two things stay constant: one material for the whole piece (sound-palette) and a deliberate relationship to silence — true silence in the dry register, silence as punctuation in the bed register (launch-video-sound).

The cue sheet

Write the sound as data next to the timeline, one row per visual event, before rendering any audio:

{ "canvas": { "w": 1920, "h": 1080 }, "fps": 30, "duration": 12,
  "bed": { "root": 43.1, "level": -22, "pad": -31, "in": 0, "out": 11.55,
           "gainPoints": [ [0, -9], [1.0, -5], [2.9, -2], [5.9, 0], [11.4, 1] ],
           "dropouts": [ { "t": 6.95, "dur": 0.6, "keep": "pad" } ],
           "swells":   [ { "t": 11.2, "dur": 0.35, "db": 2.5 } ] },
  "cues": [
    { "id": "s1-glyphs", "kind": "flicker", "t": 0.20, "n": 3,  "x": 810, "y": 480, "w": 300, "h": 132 },
    { "id": "s1-title",  "kind": "thud",    "t": 0.90, "x": 810, "y": 480, "w": 1300, "h": 132 },
    { "id": "s1-sub",    "kind": "type",    "t": 1.30, "n": 13, "every": 0.11, "x": 700, "y": 640, "w": 120, "h": 40, "gain": -8 },
    { "id": "s3-cells",  "kind": "type",    "t": 6.25, "n": 9,  "every": 0.21, "x": 604, "y": 640, "w": 260, "h": 60 },
    { "id": "s3-modal",  "kind": "thud",    "t": 7.51, "x": 1432, "y": 620, "w": 520, "h": 420, "semitones": 7 } ] }

t is the contact frame: tween start + ~85 % of its duration for a decelerating ease (where power3.out has visibly stopped), not the tween start and not its mathematical end — see sound-motion-sync. For whoosh and air, t is the tween start and dur its length; the peak lands on t + apex · dur, with apex the fraction of the tween where the move is fastest (default 1, the cut at the end of an accelerating exit; 0.5 for an ease-in-out cross; ~0.15 for a decelerating entrance). riser and zip end on t + dur, which is the cut. For type and flicker, t is the first step and the run expands into n clicks at every. A type run should not be a metronome: HeyGen's claude-paper-launch spaces its typing ticks 0.117 s apart on average (sd 34 ms, about ±30 %) with a humanizing formula, about one tick per 3 characters of text typed at 25–29 characters a second; "jitter": 0.35 on the run reproduces that spread (without it a run keeps ±8 ms, so older sheets render unchanged). Keep ticks under ~10 a second: under faster type, thin them or lay one continuous typing bed per passage, trimmed past its lead-in; --report warns on faster runs. A flicker run, streamed agent output included, is cut-exact.
kind is the gesture: click (a stepped reveal), thud (something big settles while the bed is on), type / flicker (runs), land / tick (the dry register's mallet and tick), whoosh / air (travel and overlays), success / error (a fifth or a minor second on the derived pitch), and the designed effects built to designed-sfx-anatomy's targets: riser and zip (end on the cut), impact and sting (on the reveal, tuned to the top-level key), ding runs that climb a semitone, and pop for an element appearing (double, step).
bed is the brand-film register: a drone on root (F1 by default) with an octave and a pad on the 4th, 5th, 6th, and 8th partials, a faint 0.26 s pulse, gainPoints as the film's act-by-act arc in dB, dropouts (with keep: "pad" to pull only the sub, the GPT-5 "thinking" move), swells, and automatic 3–4 dB ducking under every thud. Omit bed for the dry register.
Staggers get one cue per member with semitones stepping up — the ear hears the count.

scripts/sound-sheet.mjs cues.json --out stem.wav --report renders the stereo stem deterministically (no randomness; a fixed seed per cue id), peak-normalized to −1 dBFS (sample peak; use --peak -1.5 or check true peak against the −1 dBTP ceiling in launch-video-review), and prints the frame, pitch, length, and pan of every onset so you can check them against the timeline. --family <dir> writes six product one-shots from the same voices so the app and its launch video share a material.

Workflow

Lock picture. Cues are placed on frames; frames that move invalidate them.
Extract the events from the timeline: every from/to that translates or scales an element the viewer will notice, and every stepped set run. Fades of small labels are not events. Two settles on the same frame are one cue — voice the bigger element.
Measure the boxes. Centre and size on the composition canvas, from the layout, not from memory.
Write the arc. In the bed register, decide where the film opens low, where it fills, where it holds its breath, where it resolves — as gain points and dropouts — before touching a single click.
Render, read the report, listen once on phone speakers.
Place the stem as one clip at t = 0, inside the composition at the root, with its own id, not as seventy clips. One clip keeps sync exact. A HyperFrames product-launch-video build fetches named SFX per beat instead: leave them out under a beat track, and in registers A, B and E replace them with this stem (launch-video-sound). A mix added in post survives in one render only: HeyGen recovered three launches' soundtracks from published renders that ran 64 ms late, one frame late, and 100 ms early (commits 2a439ec, dd128ad).
Check the onset list against motion peaks the way launch-video-sound measures it: every onset on its visual event's frame or up to two frames after it, never before, and nothing in the holds.

When to apply

Logo reveals, brand films and product tours in registers A, B and E (launch-video-sound). In a beat-track sizzle the track carries the pops, and under a voice the voice is the clock; there only the track's structure is placed (launch-video-music). In product UI the same voices apply, but pan does not: UI sound is mono (sound-spec), because the element's position on a phone is not a position in the room.

Gotcha

Deriving is not the same as sonifying. If a scene has forty tweens, it does not get forty sounds; it gets the ones the eye actually tracks. In the bed register the bed is not a music track: one drone in one key, no beat, quiet enough that every click reads. In a beat or voice register the derivation moves up a level — the picture places the track's drop, dropouts and breakdowns (launch-video-music). The mapping decides how a sound behaves; sound-decision-framework and the register chosen in launch-video-sound decide whether.

Sources

HKTITAN — sound-sheet.mjs and the launch film's cue sheet in docs/demo/hyperframes/assets/sfx/cues.json (240 onsets from 138 cues, a bed with three dropouts, stem −15.4 LUFS integrated and −1.4 dBTP true peak; −18.5 LUFS in the rendered film).
OpenAI, Refreshed. and Introducing GPT-5 — the bed-and-clicks register, measured; hit-versus-sub levels and band p95 in docs/research/launch-register/*-summary.json (analysis3), the rest in launch-video-sound.
HeyGen, hyperframes-launches (Apache-2.0) — some storyboards (sfx-music-launch, hyperframes-launch, website-to-hyperframes) carry an audio cue map, but most sources shipped silent with the mix done in post; three soundtracks were recovered from published renders (commits 2a439ec, dd128ad); claude-paper-launch/index.html L168–252 for typing-tick timing (74 typing clips over 227 typed characters, 71 intervals) and claude-paper-launch/compositions/chat-response.html L325–338 for the humanizing formula.
Linear Agent (2026), Raycast (2025) — foley band, levels and harmony; Skale's Contra Indy (sub pulled 0.25–0.5 s before reveals) and Taste Labs (0.9 s at −42 dB before the raise is named). Measured by HKTITAN 2026-09-28; docs/research/launch-films/notes/synthesis-acclaimed.md, notes/synthesis-skale.md.
ITU-R BT.1359-1 — audio may lag video, never lead; see sound-motion-sync.
Blattner, Sumikawa & Greenberg, Earcons and icons (1989) — pitch, rhythm, and register as a grammar of families.
Related: launch-video-music, sound-palette, launch-video-seams, stagger-choreography.