Skip to content

Ears for coding agents: let AI hear, review and make sound. CPU-only, open source.

21.9 KB
Raw Blame

What professional motion-design SFX are made of

A measured study of the sound design in eight professional launch / motion films (third-party, kept local, called ref_1 .. ref_8 here), done to give earshot forge concrete targets: what a "pro designed" whoosh, riser, impact, click, pop or sting actually looks like in numbers, and how it sits against music and picture.

Tools: eval/forge/pro/ (earshot timeline at detail high + sheet per film, paged spectrogram overviews, zoomed crops, bed-subtracted measurements, trims). Local data: out/refs_pro/ (never committed: trims, manifest.json, registry.json, stats.json, films.json, sheets and crops under _work/).

How to read the numbers (and how far to trust them)

  • 58 events were picked by looking: earshot sheets, 8 s overview pages, then zoomed crops of every candidate. Only events that stand out over the bed were kept (an exposed accent, or an SFX-only stretch).
  • One film dominates. ref_2 is a pure sound-design piece with no music bed; it supplies 39 of the 58 events and most of the clean ones. The other seven are music- or voice-led, and their effects are sparse or buried. The per-category numbers below are often 1-10 examples from 1-2 films: they are measurements of real pro work, not population statistics.
  • Bed subtraction. Each event is measured as (event power - bed power), with the bed taken as the quietest 150 ms in the 0.6 s before the event (measure.quiet_bed, so a run of clicks or a previous hit does not count as bed). Bands where the event adds under 1 dB are marked masked and left out of the spectral medians. Over a music bed this removes most of the music, not all of it.
  • Levels are K-weighted (BS.1770 filter): over_bed = loudest 100 ms of the event minus the quiet bed. Because the bed is the quietest stretch between beats, the figure reads a few dB higher than "over the average music level". A silent bed counts as -70 LUFS.
  • Balance = bed-subtracted power per 1/3-octave band in dB relative to the event's own total (level never decides). tilt = least-squares slope of that balance from 100 Hz to 10 kHz, in dB per octave.
  • Attack = from the first millisecond within 20 dB of the peak to the peak (1 ms bed-subtracted envelope). t20 / t40 = peak to the first fall 20 / 40 dB under it (10 ms smoothing). Decay shape: convex = holds, then drops (a gate or a fade), exp = straight in dB, concave = drops fast, then a long tail.
  • Pitch = the dominant spectral ridge (or the centroid for noise). Where a film's tonal music sat under a zip, the automatic ridge locked onto the music, and the sweep was read off the zoomed spectrogram instead (marked read).
  • All audio is the films' decoded AAC at 48 kHz stereo.

The films: music first, effects placed

film dur s LUFS LRA cuts/s beat voice-over cuts with an onset within 1 frame (chance) strong accent within 1 frame (chance)
ref_1 20.5 -10.4 6.4 0.78 107 BPM, weak yes 31% (35%) 6% (8%)
ref_2 29.5 -22.2 12.2 0.54 none no 50% (44%) 19% (10%)
ref_3 49.3 -17.4 5.5 0.63 133 BPM, weak yes 58% (46%) 16% (9%)
ref_4 8.7 -23.1 0.7 1.71 111 BPM, clear no 20% (20%) 7% (4%)
ref_5 28.9 -14.8 3.6 0.31 none yes 67% (46%) 22% (9%)
ref_6 32.3 -14.8 7.0 1.27 122 BPM, clear no 81% (71%) 22% (15%)
ref_7 41.8 -7.3 3.3 1.60 92 BPM no 33% (31%) 4% (7%)
ref_8 93.1 -12.1 1.7 0.30 none yes 71% (60%) 25% (11%)

What this says:

  • Six of eight films let the music (and the voice) carry the cut rhythm. Their cuts have some onset within a frame 20-81% of the time against a chance level of 20-71% (a few points above chance; ref_5 is the clearest at 67% vs 46%), and a strong accent within a frame only 4-25% of the time. Designed effects show up at a few chosen moments: opening UI ticks over a near-silent bed (ref_6), tonal zips into cuts (ref_5, ref_6), a noise riser at the very end (ref_1), a sub drop under a cut and an end chord (ref_8). ref_4 (a cut-down) has no designed effects at all; ref_3 and ref_7 bury theirs under voice or a -7 LUFS wall of music.
  • ref_2 is the other model: a film scored entirely with effects. No beat, 12 stop-downs (drops to silence between ideas), LRA 12 LU, integrated -22 LUFS. Every effect is exposed, and every one is short, band-limited and shaped (see below).
  • So "pro" is mostly restraint and placement: few sounds, each with one clear gesture (a pitch move, a gate, a thump), cut on the picture, sitting about 13-17 dB over the music when there is music.

Per category

Counts are measured events (exported trims in brackets: forge / bench). Ranges are min-max, "med" = median.

Whoosh / swell into a cut (whoosh_big), 5 (2 / 0)

  • Made of: in the SFX-only film a whoosh is a low swell, not air: energy almost all under 150 Hz (bed-subtracted octave balance, re total: 31 Hz -19, 63 Hz -14, 125 Hz -14, 250 Hz -19, 1 kHz -28, 4 kHz -31, 16 kHz -37 dB), centroid 51-150 Hz, tilt -3 to -13 dB/oct, a faint high air layer on top at -20..-30 dB. The two buried whooshes in ref_1 are brighter (centroid 0.5-0.9 kHz) but the music makes that uncertain.
  • Envelope: 0.5-1.0 s, symmetric to slightly front-loaded: rise (-20 dB -> peak) 190 ms, fall to -20 dB 90 ms, to -40 dB 240 ms; rise shape straight-to-concave in dB (it accelerates into the peak).
  • Pitch: the centroid falls about 2.3 octaves across the pass (8.28-9.30 in ref_2): a pass-by, not a static noise.
  • Stereo: mono or near-mono (side 30-40 dB under mid); no pan travel in any example. The width in these films comes from textures, not from whooshes.
  • Timing: peak 127 ms after the cut (ref_2 8.64), or a swell that peaks 313 ms before the cut and dies into it (ref_2 2.88); the buried one in ref_1 peaks 100 ms before its cut. So: peak within about +-150 ms of the cut, with the energy straddling it.
  • Level over music: +15 to +21 dB (ref_1, two examples).

Riser (riser), 3 (1 / 1; one music riser measured only)

  • Designed riser (ref_2, 0.84 s): three layers you can see: (1) a stutter gate, 34 pulses at 22 ms median spacing (45/s), speeding up from 24 ms to 20 ms toward the cut; (2) a rising tonal layer, about 2 kHz -> 8 kHz (+2 oct; the last 0.22 s climbs 4.2 -> 8.1 kHz, about +4 oct/s); (3) a steady low drone (50-500 Hz) under it all. It ends on the cut (+20 ms) and hands over to a low whoosh / boom that blooms just after.
  • Noise riser (ref_1, 1.15 s, film end): high band (> 4 kHz) climbs about 35 dB, centroid +1.4 oct (360 Hz -> 1.1 kHz ridge), rise concave in dB (slow start, fast finish), side 14 dB under mid.
  • Music riser (ref_7, measured only): high band +20 dB over 0.45 s into the drop, only +5 dB over the music.
  • Rise shape: concave in all three: the level accelerates; most of the rise happens in the last third.
  • Level over music: +15 dB (ref_1).

Impact (impact), 3 (2 / 0)

  • Made of: sub-dominated: 31 Hz -15, 63 Hz -9, 125 Hz -12, 250 Hz -18, 500 Hz -36, 1 kHz -45 dB re total; under 150 Hz holds all but 0.7 dB of the energy in ref_2 4.71; tilt -3.6 to -5.7 dB/oct; ridge at 82 Hz.
  • Layers: a broadband transient on top (full-band vertical line at the onset), the sub body, and a textured tail: ref_2 4.71 has a granular low rumble (pulses every 28 ms, 36/s, slowing slightly) lasting 1.2 s; ref_2 14.02 has a sustained high air band (5-10 kHz) over the body.
  • Envelope: attack 50-290 ms to the loudest point (they swell into the body rather than peak on the transient); tail: t40 230 ms to 1.08 s, decay 38 dB/s for the rumble; both designed impacts end with a hard gate (the tail is cut, not left to ring).
  • Stereo: the body is centred, the rumble / air tails are decorrelated (correlation -0.01 and 0.24, side about level with mid, pan travel 28-30 dB).
  • Timing: the swell starts 110-200 ms before the cut; the loudest point lands 65 ms before to 82 ms after it.

Sub drop (sub_drop), 1 (0 / 1)

  • ref_8 85.58: a sub tone 64 -> 44 Hz (-0.55 oct over 0.75 s) entering on a cut with a bright transient on top; energy share under 150 Hz -0.5 dB of total; rise to the sub peak 430 ms (it blooms after the cut: peak +380 ms), decay 42 dB/s (straight in dB), t20 360 ms; narrow (side 13.5 dB under mid); +13 dB over the voice + music.

Logo / end sting (logo_sting), 2 (1 / 1)

  • Designed (ref_2 28.62): a falling chirp 1.5 kHz -> 150 Hz in 85 ms (-3.3 oct, about -38 oct/s) landing on a low tonal hum (fundamental about 64-80 Hz with harmonics to 160 Hz, sinking -0.6 oct) that rings about 0.8 s; mono (correlation 1.0); balance peaks at 250-500 Hz (-13 / -6 dB re total), tilt -6.9 dB/oct; t20 58 ms, t40 260 ms with a concave tail. It starts 28 ms after the cut that ends the glitch.
  • Music ending (ref_8): a struck end chord with a 4 s ringing tail (centroid 1.9 kHz, decays about 130 dB/s at first, then the tail).

Glitch / stutter (glitch), 1 (1 / 0)

  • ref_2 27.90-28.60: crackle first (sparse high grains, 4-12 kHz), then a dense broadband stutter: 11 bursts at 28 ms median spacing (36/s), speeding up from 32 ms to 22 ms into the cut; bright (centroid 9.2 kHz, tilt +3.8 dB/oct, 8 kHz band strongest); fully decorrelated (correlation 0.04, side as loud as mid, pan travel 9 dB). It ends exactly on the cut (0 ms) and the sting takes over.

Shimmer (shimmer), 2 (2 / 0)

  • Granular sparkle (ref_2 6.15, 1.95 s): dense short partial grains every 20 ms (50/s) between 1 and 10 kHz over one steady tone at about 500 Hz; swells 0.5 s, fades about 0.8 s; balance 500 Hz -11, 1 kHz -15, 2 kHz -9, 4 kHz -15, 8 kHz -20, 16 kHz -36 dB; tilt +6.5 dB/oct; wide (correlation 0.08, side 0.3 dB under mid, pan travel 14 dB).
  • Melodic chime run (ref_2 24.25, 1.95 s): glockenspiel-like notes (270 Hz - 10 kHz partials) panned note by note (pan travel 94 dB = hard-panned notes), cut off by a gate at the end.

Notification ding (notify), 4 (4 / 0)

  • A struck tone with one inharmonic partial: fundamental about 1.8-2.2 kHz plus a partial near 3.4x (6.3 kHz), attack 1-15 ms, t20 27-42 ms, t40 30-180 ms (short, convex: they ring briefly and stop), 0.23-0.36 s in all.
  • Successive dings step up a semitone: fundamentals 1822, 1934, 2045, 2168 Hz (ratios 1.061, 1.057, 1.060).
  • Balance: 1 kHz -6, 2 kHz -5 dB, nothing under 500 Hz (< -38 dB); stereo moderate (side 5.5-13 dB under mid).
  • +16 to +17 dB over the bed they ride on.

Pop (pop), 3 (2 / 1)

  • Chirp pop (ref_2 13.10): a sine falling 9.8 kHz -> 350 Hz in 110 ms (-4.8 oct, about -43 oct/s), preceded by a 20 ms blip at 3 kHz; attack 32 ms (the loudest point is mid-sweep, around 1-6 kHz), t20 14 ms after it.
  • Blip pop (ref_2 13.30): a 1.2 kHz tone with harmonics, gated on and off (70 ms box, 2 ms attack, 7 ms release).
  • Over music (ref_5 16.84): a tonal blip falling 1.1 kHz -> 100 Hz with a click, +14 dB over the music.
  • All three are mono (correlation 1.0 or side 40 dB down).

Element pop-in (pop_in, not a forge category yet), 10 (10 / 0)

  • The most used effect in ref_2: one per element appearing, fired in a run (17 events over 3.3 s, spacing 120-310 ms, median about 205 ms, settling into a steady 200 ms pulse).
  • Made of: a thump body at 80-150 Hz (ridge peak about 105 Hz; under 150 Hz holds all but 0.5 dB of the energy; tilt -4 to -6 dB/oct) plus a short noise tick (1-10 kHz, about 30 ms, 30-40 dB under the body), often as a quick double hit (two bodies 60-70 ms apart).
  • Envelope: 0.14-0.27 s, attack 6-19 ms (median 16), t20 16-35 ms (median 20), t40 30-80 ms; convex (holds, then stops) in 7 of 10.
  • Stereo: moderately wide (side 3.5-9 dB under mid, correlation 0.38-0.78).
  • Timing: lands 25-40 ms after its cut (first and last of the run).

Tap (tap), 1 (1 / 0)

  • ref_2 18.70: a 150 Hz thump (ridge 130-160 Hz) with a faint broadband click, 0.1 s, attack 11 ms, t20 18 ms, t40 27 ms, tilt -6.5 dB/oct, side 5 dB under mid; 86 ms before its cut.

Click (click), 5 (1 / 2; two measured only)

  • Bright UI clicks (ref_2, ref_6, ref_8): attack 0-3 ms, t20 6-16 ms; centroid 5-10 kHz, tilt +4 to +12 dB/oct; the ref_2 click has tonal partials at 5-8 kHz (a "tink" rather than a noise snap).
  • Mouse-like click over music (ref_5): broadband with a low body (centroid 430 Hz, tilt -3.7 dB/oct), attack 10 ms, t20 11 ms, t40 37 ms, +13 dB over the music.
  • Mostly mono (side 10-14 dB under mid). ref_6's opening click lands 31 ms before its cut.

Hover tick (hover_tick), 6 (3 / 3)

  • Tiny high ticks: 35-80 ms windows, attack 0-5 ms, t20 6-25 ms; centroid 2.8-9.5 kHz; tilt +2.4 to +6.2 dB/oct; balance: 1 kHz -23, 2 kHz -18, 4 kHz -13, 8 kHz -8 dB re total, almost nothing under 500 Hz.
  • Two kinds: tonal pings (ref_2: a single partial at 6-15 kHz, about 10 ms) and filtered noise ticks (ref_6: 2-12 kHz noise, 25-40 ms, the spectral peak falling as it decays).
  • Over a quiet bed: +15 to +20 dB.

Typing (typing_key, typing_run), 7 keys (6 / 1) + 1 run

  • Designed typing is high-passed noise, nothing else: each key a noise burst at 3-12 kHz (ridge 8.6-11 kHz), essentially nothing under 2 kHz (bands up to 1 kHz at -60 dB, the floor; 2 kHz -28, 4 kHz -13, 8 kHz -7, 16 kHz -14 dB re total), tilt +12 to +19 dB/oct; attack 2-5 ms, t20 11-18 ms, t40 about 60 ms; convex; mono.
  • The run (ref_2 10.40-11.95): keys at 62-148 ms spacing (median about 87 ms, about 11.5 keys/s), level and timing jittered, a soft low swell rising under the end of the run.
  • The bench key (ref_8, in a voice-over gap) is a more realistic mid-range key (centroid 3.2 kHz).

Zip (zip, not a forge category yet), 4 (0 / 4)

  • Fast tonal pitch sweeps (a sine-like glide plus harmonics), 0.18-0.28 s, sitting only +7 to +9 dB over the music (lower than the hits):
    • up: about 100 Hz -> 1.5 kHz in 0.2 s (read, about +20 oct/s), starting 39 ms before a cut (ref_6 1.60);
    • up: about 100-150 Hz -> 10 kHz in 0.22-0.25 s (read, about +30 oct/s), ending on the cut (+20 ms, ref_6 3.78);
    • down: 7.2 kHz -> 420 Hz in about 60 ms (read, about -60 oct/s), across a cut (peak 65 ms before it, ref_5 22.80);
    • arc: about 1 -> 3 -> 1 kHz over 0.25 s (read, ref_5 28.22).
  • Straight lines on a log-frequency axis: the sweeps are exponential (constant octaves per second).

What the pro sounds have in common

  1. One gesture each, and it is mostly pitch or gate. Zips at 20-60 oct/s, pop chirps at about -43 oct/s, the sting chirp at -38 oct/s, dings stepping up a semitone, a riser's tonal layer climbing 2 octaves. Risers and glitches are gated at 20-45 ms, and the gate speeds up toward the cut (24 -> 20 ms, 32 -> 22 ms).
  2. Band-limited on purpose. Typing has nothing under 2 kHz; pop-ins, taps, impacts and the designed whooshes live under 150-250 Hz with only a thin tick or air layer on top (20-40 dB down); dings have nothing under 500 Hz. The UI and transition sounds mostly leave 300 Hz-1 kHz (where the voice and the music's chords are) to the music; only the sting, the zips and the chirps pass through it.
  3. Hard ends. Short UI sounds hold, then stop (convex decay in 7 of 10 pop-ins, 5 of 6 keys, 4 of 4 dings); both designed impacts and the chime run end on a gate, not a natural ring-out. Long natural tails are the exception (the music end chord).
  4. Mono for objects, wide for textures. Pops, keys, clicks, whooshes and the sting are mono to near-mono; shimmer and glitch textures are decorrelated (correlation 0.04-0.36) and move (pan travel 9-94 dB); pop-ins and impact tails are in between.
  5. Placed on the cut, sound a little early. Attacks of impacts, sub drops, taps and clicks start 30-200 ms before the cut, with the body or peak landing within about 80 ms of it (sub bloom up to +380 ms); whoosh peaks within about +-150 ms; risers, the zip-up into a cut and the glitch end on the cut (0-20 ms) and hand over to the next sound; pop-ins land 25-40 ms after their cut.
  6. Levels. Over music, hits, clicks, pops and dings mostly sit +13 to +17 dB over the quiet bed between beats (K-weighted, loudest 100 ms; up to +21 dB for a whoosh), tonal zips +7 to +9 dB. In the SFX-only film, effects peak around -16 to -38 LUFS (loudest 100 ms) with silence between ideas.

Targets for synthesis recipes

Concrete numbers a recipe should hit to read as "pro designed" in this idiom (measured above; ranges = what the films do; judge with eval/forge/pro/measure.py on the render, alone, bed = silence).

category length attack tail balance pitch / motion stereo
whoosh_big (designed swell) 0.5-1.0 s -20 dB -> peak 150-250 ms, accelerating -20 dB in 80-120 ms, -40 dB in about 250 ms 80% + of energy < 150 Hz, tilt -3..-12 dB/oct, air layer 20-30 dB down centroid falls about 2 oct across the pass mono
riser 0.8-1.5 s (longer allowed) concave: half the rise in the last third ends on the hit / cut, hands over drone + tonal layer + high noise tonal layer +2 oct (2 -> 8 kHz); gate 24 -> 20 ms (40-50 pulses/s) accelerating; high band +30-35 dB over the build mostly centre (side about 14-18 dB down)
impact 0.6-1.3 s transient on top, body peaks 50-290 ms later textured tail (grains every 25-35 ms) then a hard gate; t40 0.25-1.1 s > 75% < 150 Hz, tilt -4..-6 dB/oct, ridge about 80 Hz tail grains slow down slightly centre body, decorrelated tail
sub_drop 0.8-1.5 s 30-50 ms transient, sub blooms 0.3-0.4 s after 40 dB/s straight-in-dB decay > 85% < 150 Hz 64 -> 44 Hz (-0.5 oct over 0.75 s) narrow
logo_sting (designed) 0.8-1.5 s 70 ms chirp t40 about 260 ms, concave peak 250-500 Hz, tilt about -7 dB/oct chirp 1.5 kHz -> 150 Hz in 85 ms into a hum at 64-80 Hz sinking 0.6 oct mono
glitch 0.4-0.8 s crackle then stutter ends on the cut bright: centroid about 9 kHz, tilt about +4 bursts every 32 -> 22 ms (accelerating) fully decorrelated
shimmer 1.5-2.5 s swell 0.5 s fade 0.8 s or gate 0.5-8 kHz, tilt +2..+6.5 grains every 15-25 ms over one steady tone wide (corr < 0.4), moving
notify 0.2-0.4 s 1-15 ms t20 25-45 ms, convex (stops) fundamental 1.8-2.2 kHz + inharmonic partial near 3.4x; nothing < 500 Hz a sequence steps up 1 semitone per ding near mono
pop (chirp) 0.1-0.15 s loudest mid-sweep t20 about 15 ms after it mid-high sine falling about 5 oct in 110 ms (about -43 oct/s), optional 20 ms pre-blip mono
pop (blip) 50-80 ms 2 ms 5-10 ms release (box) 1-1.5 kHz tone + harmonics none mono
pop_in 0.14-0.25 s 6-20 ms t20 about 20 ms, t40 30-80 ms, convex thump 80-150 Hz, > 85% < 150 Hz; tick 1-10 kHz 30-40 dB down optional double hit 60-70 ms apart; runs every about 200 ms slightly wide (corr 0.4-0.8)
tap about 0.1 s about 10 ms t20 about 18 ms, t40 about 27 ms 130-160 Hz thump + faint click, tilt about -6.5 none near mono
click (UI) 30-60 ms 0-3 ms t20 6-16 ms centroid 5-10 kHz, tilt +4..+12, tonal partials 5-8 kHz allowed none mono
hover_tick 20-40 ms 0-5 ms t20 6-25 ms centroid 3-9.5 kHz, < 500 Hz empty noise tick with a falling peak, or a single 6-15 kHz ping mono
typing_key (designed) 50-80 ms 2-5 ms t20 11-18 ms, t40 about 60 ms, convex noise 3-12 kHz only: < 2 kHz at least 30 dB down, tilt +12..+19 none; a run jitters spacing 60-150 ms (about 11 keys/s) and level mono
zip 0.15-0.3 s 20-35 ms ends on the cut tonal glide + a few harmonics exponential sweep 20-60 oct/s: up about 100 Hz -> 1.5-10 kHz, down about 7 kHz -> 400 Hz, or an arc near mono

Placement targets (for anything that scores a film): attacks 30-200 ms before the cut with the peak within about 80 ms of it; whoosh peaks within +-150 ms; risers / zips-up / glitches end on the cut (0-20 ms) and hand straight over to the next sound; over music, accents at +13..+17 dB over the quiet bed and tonal sweeps at +7..+9 dB; leave silence or the music alone between ideas.

The reference sets this produced (local only)

out/refs_pro/registry.json is in the earshot.forge.refsets format ({'version', 'categories': {cat: {forge, bench, note}}}, paths relative to the file), so EARSHOT_REFSETS=out/refs_pro/registry.json (or refsets.load(path)) reads it directly. The split is by film: ref_2, ref_3 and ref_7 are forge for every category; ref_1, ref_4, ref_5, ref_6 and ref_8 are bench (checked with refsets.check_split). In practice every forge trim comes from ref_2 (ref_3 and ref_7 gave nothing clean enough to export) and the bench trims come from ref_1, ref_5, ref_6 and ref_8. 51 trims (37 forge, 14 bench), trim + fade only, original gain, stereo 48 kHz 24-bit, with the music bed still underneath where there was one (each manifest entry carries its bed level, over-bed and SNR, the earshot timeline events inside the window and the nearby cuts, and every measurement).

category forge bench note
click 1 2
hover_tick 3 3
typing_key 6 1
pop 2 1
riser 1 1
logo_sting 1 1
whoosh_big, impact, shimmer, notify, glitch, tap 2, 2, 2, 4, 1, 1 0 only the SFX-only film has them exposed
sub_drop 0 1
pop_in, typing_run (new) 10, 1 0
zip (new) 0 4

The benchmark can only use categories with bench trims (click, hover_tick, typing_key, pop, riser, logo_sting, sub_drop, zip); the rest are forge-side targets. whoosh_fast, swish, swipe, toggle and downlifter were not found exposed in any of the eight films.

Rebuild: python eval/forge/pro/run_timeline.py ref_1 ... (earshot timeline + sheet), viewer.py overview|crops, study.py (measure, export, manifest, registry, stats, category sheets), extra.py (ridges, pulse rates), film_table.py. The curated event list with every time and note is eval/forge/pro/events.py.