Airwindows outboard implementation
This page documents Outboard Channel, Console Colour, Console EQ and Butter Comp: source pins, deliberate adaptations and control mappings. See airwindows-dynamics.md for Digital Black Gate.
Sources are the double-precision LinuxVST paths of CStrip2, BussColors4 and ButterComp2 at e68f75154ed0520fc5757fec9e7792dd1747e5ee. The SDK-free headers and MIT license live in dsp/airwindows/outboard/. Reproduce them with python scripts/vendor-airwindows-outboard.py; select cores with --plugins, or use --source-dir with cached upstream files. The existing reverb generator is unchanged.
Critical decisions
CStrip2 and ButterComp2 are related, different compressors. CStrip2 uses 1 + 35 H^4 input gain, a user-controlled recovery coefficient and golden-ratio sample averaging in its detector. ButterComp2 uses 10^(14 A / 20) input gain, 0.012 A / 135 / (1 + abs(lastOutput)) recovery, different automatic output compensation and no such averaging. Keep both upstream designs; do not promise that the strip is ButterComp2 with filters. Both retain separate L/R detectors.
Native Capacitor percentages are more honest than invented cutoff Hz. Both cuts expose 0–100% rather than assigning an unmeasured single corner to a rotating, three-stage ladder. Low cut 0 is off; High cut 100 is open. High cut decreases as the control is turned down. The native coefficient is p²; the port uses 1 - (1 - p²)^(44100 / hostRate) to preserve pole time constants, with both endpoints exact. Rotation remains sample-clocked, so this is compensation rather than a claim of identical transfer functions at every rate. No RBJ replacement.
Console EQ's frequency numbers are coefficient scales. Preserve the DSP ranges: Treble 1000–16000 Hz and Bass 300–16000 Hz. Their inverse mappings are sqrt((f - 1000) / 15000) and sqrt((f - 300) / 15700). They do not describe measured shelf corners. The band controls retain the DSP's nominal ±6 dB range; their sine/cosine shaping makes the actual gain level-dependent. Flat with cuts open is exactly dry, including digital silence. Bass defaults to 300 rather than upstream's surprising 4225 coefficient scale, so the first Bass adjustment works in the useful lower part of its native range. Treble retains upstream's 4750 scale. The display measures the strip's quiet-signal response from an impulse through its own DSP and draws it over the live output spectrum; louder input adds band density that this linear curve leaves out.
Outboard Channel retains the native unity-to-Spiral section. EQ is flat; Capacitor precedes the native compressor. Saturate maps to J = (1 + s) / 3, so its range drives 1–2× into the original Spiral crossfade. The Density third is omitted. Output is a separate conventional gain after Mix. Saturate changes drive as well as colour; it is not a constant-loudness blend. At Compress 0, Saturate 0, cuts off/open and Output 0 dB, nonzero input is unchanged.
Console Colour trims use actual dB. BussColors4's upstream numeric trims use /14, not /20; carrying those numbers across as dB would be misleading. Its native trims stay at centre while the wrapper Input and Output controls each cover ±18 real dB. Both affect only the wet path, so Mix 0 is bit-transparent at any trim. The eight upstream voices and built-in per-voice gains are retained, with additional fixed compensation for default-level comparison (see calibration below). Desk changes fade to dry over 10 ms, reset the small FIR/sag history, and fade back over 10 ms. They allocate no memory. This prevents old-desk convolution history leaking into the new desk.
ButterComp2 noise is local and deterministic. Its original function-static residue counters share state between instances and use signed-overflow integer squaring. The port instead uses the core's local seeded PRNG for denormal protection, matching the newer Airwindows cores. This deliberate tiny-noise deviation removes shared mutable state and undefined arithmetic. The compressor math is retained. Output maps to −18…+6 real dB after Mix, rather than the native 0…2 linear output multiplier; the internal output remains unity so the native output-dependent recovery follows the compressed signal.
Defaults and feedback
| Module | Insertion settings | Display |
|---|---|---|
| Outboard Channel | Low cut 0%; High cut 100%; Compress 20%; Comp speed 25%; Saturate 15%; Output 0 dB; Mix 100% | Measured cut response over the output spectrum and detector gain-reduction history |
| Console Colour | Dark desk; Input and Output 0 dB; Mix 100% | Desk name and live stereo peak IN/OUT histories |
| Console EQ | Bands 0; Treble scale 4750 Hz; Bass scale 300 Hz; High cut 100%; Low cut 0% | Measured quiet-signal response over the live output spectrum |
| Butter Comp | Compress 35%; Output −1 dB; Mix 100% | Detector gain-reduction history beside live peak IN/OUT histories |
Compressor percentages are intensity/speed, not threshold, ratio, or milliseconds. The GR tap reads the native per-sample multiplier, blended by Mix, before automatic gain recovery, saturation and user output. The larger L/R attenuation is shown. IN/OUT histories include the complete effect, automatic gain recovery and output. They must not be labelled "no makeup". Colour's dynamic convolution cannot be represented by a static waveshaper curve.
Continuous controls smooth over 20 ms. Core coefficients update every 32 samples using a persistent sample counter; host block partitioning does not affect that clock. The selected stereo core is allocated only during preparation. All four report zero latency. Bypass drains silence for one second before suspension; very slow native detector/filter state can persist. Menu membership follows the player decision: Distortion for Colour, EQ / Filter for Console EQ, Dynamics for the strip and Butter Comp.
Host automation and saved-state compatibility are described in development.md.
Verification and measurement
OutboardTests.h exercises five host rates (44.1/48/88.2/96/192 kHz), continuous extremes, every desk, variable blocks, exact dry paths, independent stereo and zero allocations/deletions in processing. CLAP tests check host choice metadata, automation and state recall at both ends of the slot bank. The catalog and presentation tests automatically include all four modules.
Build and run nam-outboard-audition for a reproducible 128-frame stereo workload and stimulus-specific RMS gain/GR. It includes signal generation and analysis, uses the best of two measured runs after warm-up, and excludes plugin/host overhead. Console EQ is measured with all three bands engaged, rather than its flat insertion setting. Measurements and guitar listening are separate checks.
Desk calibration
The fixed compensation in Dark/Rock/Lush/Vibe/Holo/Punch/Steel/Tube order is +0.8841, +0.6937, −1.3250, +1.0910, +0.5984, +0.7611, +2.7876, +1.1550 dB. It cancels each voice's 48 kHz offset for the tool's plucked stereo harmonic stimulus. All eight remain within 0.08 dB of unity for that stimulus across all five rates. This does not flatten their spectral differences or guarantee equal perceived loudness on other material.