Comp76Fx

Three circuit models of the classic 1966 FET limiting amplifier, one per revision, built with NIH-plug. Each builds as CLAP and VST3.

Plugin	Circuit	Character
Comp76Fx Rev A	no low noise circuit, Class A output, painted meter surround	the most aggressive and the noisiest, and its ratios undershoot their markings
Comp76Fx Rev D	low noise circuitry, Class A output	the one most people mean; the reference reissue is patterned on the D and E
Comp76Fx Rev F	low noise, push-pull Class AB output	cleaner and tighter, with the distortion turning symmetrical

They are not affiliated with, endorsed by, or connected to Universal Audio or any other manufacturer. Model numbers are used only to say which circuit is modelled.

What it models

The unit is a feedback compressor: the sidechain samples the signal after the gain element rather than before it. That is not a detail. With k as the sidechain's gain, the loop settles at

g = -k / (1 + k) * (input - threshold)

so the slope is 1 / (1 + k) and the ratio is simply 1 + k. The 4:1 button is k = 3; the 20:1 button is k = 19. The soft knee everyone describes is not dialled in anywhere, it is what a feedback loop does.

• The gain element is a FET used as a voltage controlled resistor, and it distorts the audio passing through it increasingly as it is pulled down. That is most of why the unit sounds the way it does when it is working hard.
• The recovery runs two stages together, so it is program dependent rather than a fixed curve.
• The sidechain runs out of rail rather than clipping, but only where a real one would. At equilibrium the demand and the gain reduction are the same number, so a curve that bends from the origin bends the static ratio with it and every button reads low. It stays linear across the range the unit works in and turns over only near the rail.
• All-button mode opens the knee out instead of switching to another ratio. The shifted bias leaves the sidechain with no definite point at which it starts working, so the gain arrives over a range of level and the front of a transient is through before much of it has. Because the detector is fed the compressed output, the width of that knee sets the ratio as well as the shape, which is why it lands between 12:1 and 20:1 and keeps climbing the harder the unit is driven.

Measured by the tests in core/tests/compression.rs:

4:1 button	4.02:1
8:1 button	8.07:1
12:1 button	12.12:1
20:1 button	20.23:1
all four in	15.5:1 driven, 13.1:1 gently
no buttons in	exactly 1:1, colour with no gain reduction
attack	19.5 µs to 794 µs against a marked 20 µs to 800 µs
release	50.0 ms to 1100 ms against a marked 50 ms to 1.1 s
distortion, idle at −18 dBFS	Rev A 0.48 %, Rev D 0.33 %, Rev F 0.05 %
frequency response	within 0.53 dB across 20 Hz to 20 kHz
signal to noise	97 dB, 107 dB and 109 dB

core/tests/calibration.rs holds the published figures to those tolerances, including the response at every sample rate and oversampling setting, and the meter's needle against the marks printed on its own face.

Two of those are worth reading twice. All-button mode has no single ratio: its knee is wide enough that the slope is still opening out at light reduction, so the figure only means anything at a stated operating point, which is the same caveat the manual's own "somewhere between 12:1 and 20:1" carries. And the ratios sit about a percent high because the unit's own distortion takes energy out of the fundamental the measurement reads; the loop itself, linearised, lands a percent low, and the two nearly cancel.

Controls

The panel is the hardware's. Input drives the signal against a fixed operating point, which is how the unit is threshold-less; output is make-up. Attack and release are marked slowest to fastest, which is backwards from most compressors and is how the originals were engraved.

The four ratio switches are mechanically interlocked, so clicking one releases the others. Hold shift or ctrl to latch, which is how you get all four in at once without having to be quick with your fingers.

The meter switch selects gain reduction, output level referenced to +4 or +8, or off.

The strip above the panel is not on the hardware. It carries the preset drop down, a save button and the settings button, which holds the window scale (50 % to 200 %), the oversampling quality and the dry blend.

Saved presets are one JSON file each, under a folder of the revision's own so the three do not share, and each carries a cross to delete it that asks before removing the file. A built-in preset has no file, so it cannot be deleted, and saving under its name writes a preset of your own beside it rather than replacing it in the list; replacing it would put it out of reach for good.

Built-in presets include All Buttons In: all four switches in, both dials wide open, driven hard, with the make-up set so it comes back at unity.

Building

./install.sh

Builds all three and installs the CLAP and VST3 of each into ~/.clap/BurningTreeC and ~/.vst3/BurningTreeC. Pass --no-build to install what is already built, or set CLAP_PATH and VST3_PATH to install elsewhere.

To build without installing:

cargo xtask bundle -p comp76fx_rev_a -p comp76fx_rev_d -p comp76fx_rev_f --release

To try one without a host:

cargo run --release -p comp76fx_rev_d --features standalone -- --backend auto

Licensing

Under the GNU General Public License version 3 or later, whose text is in LICENSE. NIH-plug itself is ISC licensed, but nih_export_vst3!() links the GPLv3 vst3-sys bindings, so any VST3 built with it has to be able to comply with the GPL.

THIRD-PARTY-NOTICES.md reproduces the licences and copyright notices of every crate the plugins link. Noto Sans is compiled in for the panel lettering and is under the SIL Open Font License 1.1. Regenerate the file after changing dependencies:

python3 tools/third-party-notices.py

Layout

Path
core/src/dsp/detector.rs	the feedback sidechain and its timing
core/src/dsp/fet.rs	the gain element and its distortion
core/src/dsp/amp.rs	the Class A and Class AB output stages
core/src/editor/	the front panel
core/src/presets.rs	built-in and saved presets
rev_a, rev_d, rev_f	one identity each; the circuit differences live in Revision
core/tests/compression.rs	the measurements above