Free one-knob compressor for macOS (VST3/AU) that finds its own sweet spot — and, turned all the way up, stops being polite about it.
The knob has two halves, and they behave differently on purpose.
Bottom two thirds — a compressor that leaves the loudness alone.
Most compressors need setting up before the knob does anything useful: park the threshold at a fixed dBFS value and a quiet take sits below it doing nothing while a hot one slams. This one measures how loud the material is running and places its threshold relative to that, so a quietly recorded vocal and a hot one get roughly the same amount of compression at the same knob position — measured, within about 1.5 dB across a 15 dB range of source level. You can start turning without setting anything up first.
And whatever loudness it takes out, it puts back automatically — not estimated from the threshold and ratio, but measured, as K-weighted energy either side of the compressor. So turning the knob up here does not make anything louder. The dynamics tighten, the sound gets denser and steadier, and the level stays where it was.
Top third — the wall.
From knob 24 up it stops restoring the level and starts deliberately driving into its own brickwall limiter: the threshold goes deeper, the detector window shortens, and the release lengthens so the gain holds instead of springing back between hits. The result is dense, loud and sitting on the ceiling. On a drum loop the short-term level swing goes from 14.2 dB in to 5.6 dB out with the peaks on the ceiling. The readout says CRUSHED for a reason.
Below knob 24 none of that is active, so the gentle range is untouched by it.
The ATTACK knob in the ADV panel decides how much of a transient survives this. It moves two things at once, because in this design they say the same thing: the envelope's attack, and how much of the 1.5 ms lookahead is spent pre-empting the hit. Driving only one gives half a control — measured, each is worth about 4 dB on its own and they overlap almost completely. Reported latency does not move either way.
RELEASE, beside it, is the other half of the same idea: Attack shapes the single hit, Release shapes the groove between them. It is set in note values rather than milliseconds, because what matters on a loop is where the gain recovery sits against the beat, and that relationship should survive a tempo change.
AUTO drives the knob to where the compressor is delivering 3 to 5 dB on peaks — the usual place for a vocal — and follows the material. Drag the knob away and it holds while the mouse is down, then pulls back like a rubber band when you let go, harder the further you dragged it. It parks well below where the wall starts, so switch AUTO off if you want to hear the top.
The sweet spot is a claim about gain reduction, not about knob position: its
edges are where the compressor actually delivers 3 and 5 dB, and auto_probe
measures both every run.
You can hear what the compressor is doing, not just that it got louder. This is the real point. On almost any compressor, turning it up makes it louder, and louder always sounds better at first — you turn it up, you like it, and the next day the track is flat. Here the level stays put through the working range, so if it sounds better as you turn up, it is because it sounds better.
No gain staging first. Drop it on the track and turn the knob. The position means roughly the same thing on every source.
AUTO finds the spot for you and keeps following it as the performance changes.
And the top third is an effect, not an accident. When you want a drum loop or a bus genuinely crushed, that is where you go. It gets a lot louder up there on purpose — which is what TRUE LEVEL is for: it takes that loudness advantage back out so you can judge what the wall is doing to the sound rather than how loud it is.
Short version: a tool at the bottom, an effect at the top, one knob.
The ingredients are not new. The combination, as far as I could establish, is.
Automatic makeup gain is everywhere — but it is almost always calculated from the threshold and ratio rather than measured. FabFilter's own Pro-C manual calls its version "an educated guess" that you may need to correct by hand; Tokyo Dawn Labs refuse to ship one at all and say why in print. A makeup that actually measures the loudness it is replacing is ahead of what the category normally does.
Loudness matching for fair A/B exists too — iZotope ships it, for one. But as a separate feature alongside the makeup, not interlocking with it.
What I could not find anywhere is both behaviours on one knob: a compressor that holds measured loudness constant through its working range, and the same knob turning into a limiter wall at the top. sonible splits exactly these two behaviours across two products — their compressor has auto gain and no match switch, their limiter has the match switch and no auto gain. And the best-known one-knob compressor in this category, Waves RVox, has no answer to the comparison problem at all: its manual suggests ganging two faders together.
So not "nobody has done this", but: the pieces are known, and assembled this way I have not found it.
In Trim → Compressor (auto-threshold, gate, 2x oversampled) → Makeup
→ TRUE LEVEL → Limiter → Dry/Wet Mix → Out Gain
The limiter's ceiling is fixed at −0.3 dBFS and it holds it to within 0.03 dB. Through the lower two thirds of the knob it does nothing at all; the wall is the only thing that drives into it. Reported latency is 158 samples at 44.1 kHz (3.58 ms), fully compensated for the dry/wet mix and for bypass.
| Control | Range | What it does |
|---|---|---|
| Compression | 0–36 | The main knob. Auto-tracking threshold and ratio; a wall above 24. |
| AUTO | on/off | Drives the knob to the sweet spot and keeps following it. |
| TRUE LEVEL | on/off | Holds output loudness equal to the input, so you can compare settings — or the plugin against bypass — without the louder one winning by being louder. The figure beside it reads in both states: engaged it is what the match is applying, disengaged it is what engaging it would cost. Below knob 24 that is a fraction of a dB, because the makeup already holds loudness there; at the top of the knob it is the wall's 14 to 16 dB. |
| Gate | −80 to −20 dB | Noise gate, off at minimum. Its threshold is drawn on the input meter, with the current attenuation next to it. |
| Attack | AUTO, 0.1–20 ms | How much of a transient survives. At AUTO the peak is caught before it arrives and nothing gets through, which is what a vocal wants. Turn it up and the grab is delayed and the lookahead pre-emption is withdrawn together, so the hit punches through before the body is squashed. Measured on a drum loop at Comp 24 it moves the output crest by 4.2 dB while the loudness stays within 0.05 dB — you are hearing character, not level. Above Comp 30 the limiter takes back what it lets through, so the knob dims there rather than pretending. |
| Release | AUTO, 1/32–1/2 | How hard the loop breathes, in note values against the session tempo — so the same setting keeps the same relationship to the groove at any BPM. Measured at Comp 24 on a 120 BPM loop, the gain swells 9.4 dB between hits at AUTO and 2.6 dB at 1/2: a loop that pumps against one that sits still. Loudness holds within 0.32 dB. Fades out above Comp 28 with the wall, like Attack. |
| SC HP Filter | 0–400 Hz | High-pass on the detector only, so bass does not pump the gain reduction. The panel draws its actual response. |
| Mix | 0–100% | Parallel compression blend, latency-compensated. |
| In Trim | ±12 dB | Input gain, ahead of the detector. |
| Out Gain | −24 to +12 dB | True output level, applied after everything. Not the limiter ceiling. |
| Bypass | — | Latency-compensated, click-free. |
AUTO and TRUE LEVEL are buttons on the interface rather than automatable parameters. Everything else is exposed to the host.
- macOS (Apple Silicon or Intel)
- A VST3 or Audio Unit host (Ableton Live, Logic Pro, etc.)
No signed release yet — see INSTALL.md for how to remove macOS's quarantine flag from an unsigned build.
Prebuilt binaries will appear under Releases once available. Until then, build from source below.
Requires CMake 3.15+ and Xcode command line tools.
git clone https://github.com/frankknebeljanssen-create/SmartComp.git
cd SmartComp
./build_and_install.shThis clones JUCE into
~/Library/Caches/SmartComp-juce (kept outside the repo and outside any
Dropbox-synced folder — a synced build directory causes CMake cache
conflicts), builds the plugin, and installs the VST3 to
~/Library/Audio/Plug-Ins/VST3/. The AU and standalone builds are produced but
not installed.
Claims about how this thing behaves are checked against the real code, because on this project reasoning about the DSP and porting it to Python have both produced wrong numbers. Three harnesses, all runnable:
tests/dsp_bench.cpp exercises the DSP classes directly — limiter ceiling,
transient response, delivered time constants, and whether a knob position
means the same thing at different source levels.
cmake --build ~/Library/Caches/SmartComp-build --target dsp_bench
~/Library/Caches/SmartComp-build/dsp_bench_artefacts/Release/dsp_benchtests/density_probe.cpp drives the whole processor and asks what comes out
over time: short-term loudness spread, how far the plugin's own gain swings,
how close the output gets to the ceiling, whether the result depends on how
hot the source was, and what the drive does to a room bed in the pauses. Two
reference rows bracket what is reachable on the material at all — the limiter
driven hard, and a perfect 20 ms AGC.
cmake --build ~/Library/Caches/SmartComp-build --target density_probe
~/Library/Caches/SmartComp-build/density_probe_artefacts/Release/density_probetests/auto_probe.cpp is a behavioural test for AUTO: that the knob converges
on the sweet spot, that the band's edges deliver the 3 and 5 dB they promise,
that a drag holds while the mouse is down and pulls back when it is released,
and that parking the knob in the red with AUTO off recovers rather than
sticking.
cmake --build ~/Library/Caches/SmartComp-build --target auto_probe
~/Library/Caches/SmartComp-build/auto_probe_artefacts/Release/auto_probeThe screenshot above is generated the same way rather than mocked up:
tests/ui_shot.cpp runs the real processor and editor through a genuine JUCE
event loop against a synthetic vocal, then saves what actually renders.
cmake --build ~/Library/Caches/SmartComp-build --target ui_shot
# at rest: AUTO on, knob snapped into the sweet spot
open ~/Library/Caches/SmartComp-build/ui_shot_artefacts/Release/ui_shot.app --args /tmp/screenshot.png
# a fixed knob position instead, with AUTO off so it stays there
open ~/Library/Caches/SmartComp-build/ui_shot_artefacts/Release/ui_shot.app --args /tmp/wall.png 36Driven hard, the wall lifts whatever is in the pauses along with the music. A room bed at −70 dBFS still sits 48 dB under the words and one at −60 dBFS 38 dB under, both comfortably out of the way. At −50 dBFS the protection cannot keep up and the gap closes to 7 dB. Gate it, or back the knob off.
Not yet decided. The repository is public for now; no license has been granted for reuse. This section will be updated before the first release.
SmartLim, a matching limiter, is planned as the second half of a set.