aesthetic.computer / core
Monorepo for Aesthetic.Computer aesthetic.computer
Star
Fork 0
Watch
core pop dsp
1 folder · 2 files
c
eq-graph.json
README.md
README.md

acdsp — Aesthetic Computer DSP #

Last updated: 2026-05-23

physically-modeled compressors and an EQ knowledge graph, written in portable C99 so the same code runs in three places:

  1. the pop/ mastering pipeline (native CLI),
  2. the AC browser runtime (wasm — pieces can call comp/EQ at paint),
  3. AC Native (static lib, when fedac/native grows audio out).

zero deps beyond libm. one library, one source of truth. no ffmpeg acompressor approximations.

why c #

ffmpeg's acompressor is a generic feed-forward VCA — clean, but program- independent envelopes and no harmonic generation. that's the audible "tell" of the current pop/ masters and the reason every lane keeps tacking a brightening pass on the end. real compressors are physical: a JFET's asymmetric transfer curve, a vactrol's two-time-constant release, a tube's soft-clip curve. modeling them needs sample-accurate state, not filter strings.

c was picked because:

  • AC Native is c — same toolchain, same calling conventions
  • c compiles trivially to wasm via emcc, no glue
  • the resulting acdsp.wasm can later become an AC piece API (comp.la2a({ peak: 8 }) inside a paint)
  • the lowest-common-denominator dependency story (libm only)

build #

cd pop/dsp/c
make            # → ./acdsp (native CLI)
make lib        # → ./libacdsp.a (link target for AC Native)
make wasm       # → ./acdsp.js + ./acdsp.wasm (needs emcc on PATH)
make test       # synthesizes a sine, runs it through the chain

status #

stage done next
native CLI ✅
wasm build target ✅ stubbed wire into AC runtime piece API
AC Native static lib ✅ via make lib actual integration when audio out lands
biquad EQ (RBJ cookbook) ✅
EQ knowledge graph ✅ codegen eq_graph.c from JSON
1176 (FET) ✅
LA-2A (vactrol) next — two-time-constant release
Fairchild 670 (vari-mu) tube transfer curve + prog. recovery
SSL G-bus (VCA + SC HPF) real sidechain HPF
loudnorm (ITU-R BS.1770) so ffmpeg shrinks to demux/encode
true-peak limiter with 4× oversample

what's physical about the 1176 #

src/comp_1176.c. not a knob-mapped acompressor:

  • stereo-linked peak detector (max of |L|, |R|) feeds a single-pole envelope smoother. coefficient is exp(-1 / (τ · sr)). attack when the static curve wants more reduction; release otherwise.
  • soft-knee static curve (RBJ-style quadratic), threshold internal at −10 dBFS with a 6 dB knee. the user gets the unit's input/output knobs; drive the input harder to push more signal into the curve.
  • FET asymmetric saturation with positive DC bias for 2nd-harmonic dominance — the audible FET fingerprint. drive depth scales with the instantaneous gain-reduction amount: 0 dB GR ⇒ 30% of base depth, 18 dB GR ⇒ 100%. the FET conducts harder ⇒ more nonlinearity. this is the part acompressor cannot do.
  • A12 line-amp tanh on the way out — gentle symmetric warmth from the post-FET stage.
  • attack/release knobs map to the unit's actual time sweep:
    • attack 1 → 800 μs, attack 7 → 20 μs (exponential)
    • release 1 → 1100 ms, release 7 → 50 ms (exponential)
  • ratios 4 / 8 / 12 / 20.

references for the physics: Eichas/Möller JFET papers (2014–2017), Zölzer's DAFX second edition (envelope/curve formulations), Airwindows 1176-ish sources (MIT, for sanity-checking sat curves). GPL sources were not consulted — LSP/Calf/Steinberg are off limits.

eq knowledge graph #

pop/dsp/eq-graph.json is the source of truth. each entry resolves to a single biquad (peak / shelf / HP / LP), carrying a frequency center, Q, default gain, a why: note, and optional conflicts: / pairs_with: hints. the C lib mirrors the table in src/eq_graph.c (kept in sync by hand for now; a codegen step is in the TODO).

intent kind freq default gain use when
sub highpass 30 Hz — strip subsonic; FM + small speakers can't reproduce
rumble highpass 60 Hz — aggressive low cut for material with no sub content
warmth peak 150 Hz +1.5 body without mud; lifts chest of vocal / weight of synths
mud peak 250 Hz −2 low-mid cloudiness from stacked synths/sub-pad
chest peak 220 Hz +2 vocal body restore (conflicts with mud)
boxy peak 500 Hz −2 cardboard-tube honk on vocals / overdriven mids
nasal peak 900 Hz −2 pinched / nasal on vocals + close-mic'd dialogue
honk peak 1200 Hz −2 upper-mid bark on saws, leads, bright synths
harshness peak 3000 Hz −2 edginess on cymbals, guitar tops, bright vocals
presence peak 4000 Hz +2 forward / in-the-room; lifts vocals through a busy mix
bite peak 5500 Hz +1.5 snap on transients — beater click, snare crack, pick
sibilance peak 7500 Hz −3 s/t harshness on vocals (narrow Q)
air highshelf 11 kHz +1.8 open top — silk on vocals, sheen on cymbals
sparkle highshelf 14 kHz +1.5 glass top; mastering-only
tilt-bright highshelf 8 kHz +1.2 broad tilt for dark mixes
tilt-warm lowshelf 200 Hz +1 broad warm tilt for thin material

adding an intent: edit the JSON, mirror the row in src/eq_graph.c, make.

chain spec syntax #

stages are colon-separated; multiple stages are space-separated.

1176:ratio=4:in=-12:out=+6:attack=4:release=4:iron=0.5

eq:<intent>[=gain_db]            # knowledge-graph lookup
eq:presence=+2
eq:air                           # uses graph's gain_default
eq:peak:f=3500:q=1.2:g=+2        # raw biquad
eq:hp:f=30                       # raw highpass
eq:highshelf:f=11000:g=+1.8

1176 keys:

key meaning range / default
ratio 1176 ratio (4 / 8 / 12 / 20) 4
in input-knob trim in dB (drives signal toward curve) 0
out output-knob make-up in dB 0
attack knob units 1..7 (1=800 μs, 7=20 μs, exponential) 4
release knob units 1..7 (1=1100 ms, 7=50 ms, exponential) 4
iron FET + A12 nonlinearity depth 0..2, default 0.5
attack-us direct override in μs (overrides attack knob) —
release-ms direct override in ms (overrides release knob) —
threshold internal threshold override (rarely needed) −10 dBFS
knee soft-knee width in dB 6

cli #

acdsp in.wav out.wav --chain "stage1 stage2 ..."
                     [--bits 16|24|32]   # PCM int width (default 24)
                     [--float]           # 32-bit IEEE float output
                     [--quiet]

example — a danceable master-bus pass:

./acdsp scratch-mix.wav master.wav \
  --chain "eq:sub 1176:ratio=4:in=-3:out=+3:iron=0.5 eq:mud=-1.5 eq:presence=+2 eq:air=+1.5" \
  --float

input WAV formats accepted: PCM 8 / 16 / 24 / 32 int, IEEE float 32, mono or stereo, any sample rate.

node api (pop/lib/master.mjs) #

import {
  acdspAvailable,    // → boolean: is the native binary built?
  eqGraph,           // → parsed eq-graph.json (with why / conflicts notes)
  compressor,        // → "1176:..." spec fragment
  eq,                // → "eq:..." spec fragment
  chain,             // → joins fragments with spaces
  processWav,        // → runs the CLI, returns { ok, ms, stderr }
  presets,           // → curated starting chains
} from "pop/lib/master.mjs";

// build a chain by knobs:
const spec = chain(
  eq("sub"),
  compressor("1176", { ratio: 4, in: -3, out: +3, attack: 5, release: 3, iron: 0.5 }),
  eq("mud", -1.5),
  eq("presence", +2),
  eq("air", +1.5),
);

// or pick a preset and tune it:
const spec2 = presets.danceMasterBus({ in_db: -3, out_db: +3, iron: 0.7 });

// run it:
const r = processWav("in.wav", "out.wav", spec, { float: true });
if (!r.ok) throw new Error(r.stderr);

presets currently shipped:

  • danceMasterBus({ in_db, out_db, iron }) — sub HP → 1176 4:1 → mud cut → presence → air
  • vocalLead({ in_db, out_db, iron }) — rumble HP → 1176 8:1 → nasal cut → presence → de-ess → air
  • drumSlam({ in_db, out_db, iron }) — 1176 20:1 ("all-buttons" feel) → bite → air

using it from a lane #

pop/dance/bin/bake.mjs is the reference pattern. opt in with --acdsp:

node pop/dance/bin/bake.mjs --acdsp

without the flag, the original ffmpeg-only chain runs (zero risk for shipping lanes). with the flag, the C 1176 + EQ chain runs before ffmpeg's loudnorm/limit/fade, between the post-FX stage and the final encode. existing lanes can be refit one file at a time — see pop/dance/bin/bake.mjs:60-95 for the diff.

how to add a new compressor #

  1. write src/comp_<name>.{h,c} — a single struct + init + update + process(buf, n_frames, channels). study comp_1176.c as the reference shape.
  2. add a parser branch in src/chain.c (mirror parse_1176).
  3. add the source file to SRC_LIB in the Makefile.
  4. make — done. it'll be callable via --chain "<name>:..." in the CLI and via the node shim.

if you're modeling a new unit, the priority is the physical fingerprint (envelope shape, harmonic content, sidechain HPF if any, program dependency), not the knob count. an LA-2A with attack/release knobs is not an LA-2A.

roadmap #

  • LA-2A: vactrol envelope (LED → LDR resistance lag), fast follower + slow leak release model. shorter code than 1176; ~60 lines.
  • Fairchild 670: vari-mu tube — gain reduction is the tube's transfer curve, so GR amount controls the harmonic flavor. 6-position recovery switch with multi-stage release.
  • SSL G-bus: clean VCA + real sidechain Linkwitz-Riley HPF + the 4K manual's auto-release curve.
  • loudnorm: ITU-R BS.1770-4 + EBU R128 integrated LUFS + LRA + TP.
  • true-peak limiter: 4× oversampled with ISR detection.

once loudnorm + limiter land, ffmpeg shrinks to a demux/encode wrapper and the DSP is entirely portable.