aed428e451
Measured at the new default threshold against a 1Vpp humbucker, as dB of dynamic range removed: ratio pot travel dB removed 1.0 0% 0.0 2.0 5% 6.1 4.8 20% 9.7 8.0 37% 10.5 20.0 100% 11.2 The first twentieth of the knob covered more than half the effect and the last 63% covered 0.7dB. Anywhere above a third of travel was indistinguishable from anywhere else above a third of travel. The cause is that the pot was linear in the wrong quantity. compressor_init() stores 1 - 1/ratio, which is what the exponent uses, and that saturates: 1->2 moves it from 0 to 0.5 while 10->20 moves it from 0.90 to 0.95. A pot linear in ratio is crowded at the bottom and flat at the top by construction, whatever range it is given. EXPONENTIAL(1 20) needs no new curve, keeps 20:1 reachable and puts the useful 1..8 span across 69% of the travel instead of 37%. The values do not change; where they sit on the knob does. This was missed when the same effect was swept an hour earlier, because that sweep asked whether each default was sensible - 4.8 is - rather than whether the travel was usable. Different questions, and the second has now caught two controls on this effect. pots.py is the other half. Analysis scripts had been converting engineering values to the 0..120 the firmware stores by writing the range out inline, and that is how three measurements in one afternoon got taken at settings nobody asked for: a range moved and the script did not, and the numbers came out plausible and wrong. It reads the POT: line instead, and checks itself - gen_effects.py has already converted every declared default into a raw value, so converting the same default here and comparing is a test of this arithmetic against the generator's, on all 65 linear and exponential pots, every run. It refuses FREQUENCY and SQUARED rather than guessing at them. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
12 KiB
Compressor [COMPRESSOR]
Five controls: Level (−60..−10 dB, the threshold), Attack (2..100 ms), Release (50..500 ms), Ratio (1..20×) and Boost (0..24 dB, makeup gain). Defaults are −35 dB, 15 ms, 150 ms, 4.8× and 6 dB.
The arithmetic is short enough to state. An envelope follower tracks the
rectified input with those attack and release times; above the threshold the
gain becomes (level/env)^(1 − 1/ratio), slewed to avoid clicks; and the
result is multiplied by Boost whatever happens. So Boost is unconditional
and compression only ever subtracts — quiet playing gets the makeup gain and
nothing else, loud playing gets the makeup gain minus whatever the ratio takes
away.
The static answer, which is exact and not very useful
A steady tone at a series of levels, measuring the gain the compressor settles on:
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00'}}}}%%
xychart-beta
title "Gain applied to a steady tone. Level -20 blue, Level -30 orange"
x-axis "input, dBFS peak" -60 --> -10
y-axis "gain, dB" -10 --> 8
line [6, 6, 6, 6, 6, 6, 6, 6, 6, 2.95, -1.01]
line [6, 6, 6, 6, 6, 6, 6, 2.95, -1.01, -4.97, -8.94]
The knee sits exactly where Level says, and above it each 5 dB of input buys 1.04 dB of output — a ratio of 4.81, against the 4.8 on the pot. Nothing is wrong with it.
It is also nearly worthless as a description of the effect, which is why this page does not stop here. A compressor's whole job is changing its gain over time in response to a signal that changes over time, and a tone that never changes measures the one case that never happens.
The dynamic answer
Inputs/BassForLinus.mp3 is a friend of mine playing scales on a bass, 77
seconds of real dynamics. It is committed unmodified, so it matches the copy
it came from.
The level it is played at is a measurement choice, not a detail. The mp3 decodes to +2.86 dBFS — reconstruction overshoot on a master that was already close to full scale — so used raw it would clip the pedal's input before the compressor saw any of it. It is scaled to −6 dBFS peak, which is where an instrument track is usually aimed, giving −25.9 dBFS RMS over the whole take. A quieter input crosses the threshold less often and everything below moves with it.
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00'}}}}%%
xychart-beta
title "Level -30: input blue, output orange, half-second windows"
x-axis "seconds" 0 --> 76
y-axis "dBFS rms" -75 --> -10
line [-69.5, -73.8, -71.6, -66.2, -25.3, -22.4, -25.1, -24.7, -23.6, -25.4, -28.2, -29.8, -33.6, -34.5, -35.3, -35.1, -32.7, -34.4, -32.4, -32.0, -30.7, -38.4, -36.3, -34.2, -36.1, -30.8, -27.9, -25.1, -23.6, -24.9, -26.0, -24.6, -28.6, -26.5, -25.6, -24.3, -24.5, -25.3, -26.5, -24.2, -25.7, -26.7, -27.4, -28.0, -36.9, -47.4, -63.5, -25.4, -24.5, -22.5, -22.1, -23.9, -25.4, -27.4, -36.1, -24.6, -24.6, -24.3, -23.7, -25.0, -27.0, -29.5, -37.7, -24.3, -24.7, -26.8, -23.3, -22.5, -24.9, -33.2, -41.3, -21.6, -27.8, -29.4, -28.4, -22.7, -23.9, -25.3, -32.2, -67.4, -62.0, -58.6, -54.3, -18.9, -23.5, -49.0, -31.7, -28.4, -23.7, -51.7, -54.0, -20.7, -22.5, -23.7, -30.9, -31.5, -42.8, -63.0, -35.9, -21.4, -22.1, -19.7, -25.8, -30.4, -39.1, -60.7, -44.6, -21.7, -23.1, -20.0, -20.5, -21.2, -23.7, -26.4, -34.6, -21.5, -22.0, -23.0, -24.1, -24.9, -25.2, -25.7, -29.0, -24.1, -25.8, -27.7, -28.3, -22.3, -23.2, -24.4, -25.1, -24.0, -24.5, -24.8, -27.5, -32.8, -29.2, -29.5, -28.2, -29.7, -28.5, -30.9, -30.2, -30.1, -27.1, -31.6, -34.2, -37.9, -25.8, -25.6, -26.3, -26.8, -27.4]
line [-63.5, -67.8, -65.6, -60.2, -24.4, -24.3, -25.2, -24.9, -24.2, -25.3, -26.3, -26.1, -28.1, -28.5, -29.3, -29.1, -27.8, -28.6, -27.5, -27.2, -27.9, -32.4, -30.3, -28.4, -30.3, -25.9, -25.6, -25.2, -24.6, -25.1, -25.4, -25, -26.5, -25.3, -25.2, -25.1, -24.8, -25.9, -26.2, -24.2, -25.5, -25.7, -25.5, -25.4, -33.3, -41.4, -57.5, -24.3, -24.5, -24, -24.3, -24.9, -24.8, -25.2, -31.8, -24.6, -24.6, -24.8, -24.6, -25.6, -25.6, -26.2, -32.9, -23.9, -24.6, -25.5, -24.1, -24.6, -25.5, -31.3, -35.3, -23.2, -26.5, -26, -26.4, -23.8, -24.2, -24.7, -31.2, -61.4, -56, -52.6, -48.3, -21.9, -28.1, -43, -27.6, -25.8, -27.4, -45.7, -48, -22.8, -26.8, -24, -27.3, -26.8, -37.2, -57, -30.2, -23.6, -25.4, -24.1, -25.9, -27.2, -33.1, -54.7, -38.6, -23.3, -26.6, -23.1, -24.7, -25, -25.8, -26.4, -32.2, -23.4, -24.6, -24.8, -25.1, -25, -24.8, -25, -27.8, -23.8, -24.8, -25.1, -25.1, -23.6, -24.2, -24.5, -24.6, -24.7, -24.8, -24.8, -27, -27.4, -26.1, -27.4, -25.8, -26.8, -26.1, -27, -27.1, -26.8, -26.1, -26.5, -28.2, -31.9, -24.9, -25.3, -25.4, -25.3, -25.2]
That is the picture a tone cannot give. The blue line wanders over 50 dB; the orange one spends most of the take between −24 and −28 dB. The passages the player left quiet come up, the loud ones are held, and the shape of the phrase survives — the orange line still rises and falls, it just does it over a quarter of the range.
The silences are the exception and they are correct: where the input drops to −67 dB between takes, the output sits 6 dB above it, because there is nothing to compress and the makeup gain is unconditional. A compressor lifts the noise floor between notes; this one is honest about it.
What the threshold is worth
The same recording, gain plotted against how loud the playing was, at two threshold settings:
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00'}}}}%%
xychart-beta
title "Gain against playing level. Level -20 blue, Level -30 orange"
x-axis "input, dBFS rms in half-second windows" -55 --> -16
y-axis "gain, dB" -6 --> 8
line [6, 6, 6, 6, 6, 5.99, 5.5, 3.7]
line [6, 6, 5.87, 5.47, 4.23, 1.34, -1.3, -3.5]
At −20 dB, against this recording at this level, the effect never applies less than +3.0 dB and never more than +6.0 — the blue line is flat until the loudest few seconds and then bends slightly, which is a boost with a hint of compression on top. The orange line, ten decibels down, is what a compressor's transfer curve is supposed to look like: +6 dB on the quiet passages falling to −3.5 dB on the loud ones, about 10 dB of range squeezed out of the performance.
Neither is right or wrong. Which one you get depends on how loud the signal arriving is, and that is a thing the player sets — with Trim, with a boost, or with whatever pedal is in front. What the threshold does is fix the distance between the two.
A caution about reading the two transfer curves together: they do not share an x-axis. The static one is the peak amplitude of a sine; this one is half-second RMS of a bass. Those differ by the crest factor, which for plucked bass is a good 12 dB, so the knee appears in a different place on each and neither is wrong.
What the knobs are worth
The threshold is compared against an envelope of the input, so how much any setting does depends entirely on how hot the instrument is. That makes "is this default any good" a question with three different answers:
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00, #009e73'}}}}%%
xychart-beta
title "Dynamic range removed. Input peaking -6 blue, -12 orange, -20 green"
x-axis "Level, dB" -40 --> -15
y-axis "dB of range removed" -1 --> 17
line [15.7, 12.0, 8.2, 4.4, 0.9, -0.0]
line [11.2, 7.4, 3.6, 0.5, -0.0, -0.0]
line [5.2, 1.5, -0.0, -0.0, -0.0, -0.0]
At a realistic instrument level the default threshold does nothing at all. A guitar around 0.1 V RMS lands near −12 dBFS peak once its crest factor is allowed for, and on that middle line the default −20 dB removes 0.0 dB of dynamic range. You have to be at −35 before it takes out a useful 7 dB, and −35 is most of the way to the bottom of a pot that stops at −40.
The top half of the Level control — everything above about −25 — does nothing for any input a guitar produces.
Attack and release cost low notes
A fast envelope follows the waveform rather than the note, and modulating the gain at the note's own frequency is distortion. It is a bass problem far more than a guitar one: the penalty falls about 11 dB per octave.
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00, #009e73'}}}}%%
xychart-beta
title "THD by note. Release 50ms blue, 150ms orange, 500ms green"
x-axis "Hz" [40, 80, 160, 320, 640]
y-axis "THD, dB" -105 --> -35
line [-40.5, -50.6, -62.0, -73.9, -85.8]
line [-48.3, -58.5, -69.9, -81.8, -93.8]
line [-57.6, -67.8, -79.2, -91.1, -103.1]
At the default 150 ms, a guitar's low E measures −58.5 dB and a bass low E −48.4 dB. So the default is comfortable for the instrument this pedal is for and marginal for the one an octave below it. Dropping to 50 ms costs 8 dB everywhere and takes bass low E to −40.5 dB, which is audible.
Attack barely matters here by comparison — the whole 2 ms to 100 ms range moves THD by 5 dB — while it does change how much gets squeezed, from 9.9 dB at 2 ms to 4.3 dB at 100 ms. So attack is close to a free control and release is not.
Ratio is nearly free above about 8: at Level −30 it buys 8.8 dB of squeeze at 8:1 and 9.4 dB at 20:1, against 8.2 dB at the default 4.8.
What the defaults are, and why
These were made up. The controls were given plausible-sounding numbers when the effect was written and had never been measured. The ranges too.
Measuring changed one of them. Because the response is a pure translation, the
end of the pot is a wall: with the old LINEAR(-40 0), a guitar arriving at
about −20 dBFS peak could not get more than roughly 5 dB of compression at any
setting, simply because the knob stopped. The range is now LINEAR(-60 -10),
which costs 0.42 dB per step against 0.33 and forecloses nothing.
With that room, here is what each candidate default does to the instrument somebody actually plugs in — no trim, nothing in front:
%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00, #009e73, #666666'}}}}%%
xychart-beta
title "dB removed by pickup. 280mVpp blue, 500mVpp orange, 1Vpp green, 2Vpp grey"
x-axis "Level, dB" [-30, -33, -35, -37, -40, -45]
y-axis "dB of range removed" -1 --> 21
line [-0, 0.4, 1.4, 2.9, 5.1, 8.9]
line [1.4, 3.5, 5.1, 6.6, 8.9, 12.7]
line [5.9, 8.1, 9.7, 11.3, 13.5, 17.0]
line [10.4, 12.7, 14.3, 15.8, 17.6, 20.5]
Peak-to-peak volts convert through process.h: a sample is the instantaneous
volts over √2, because 1.0 is a 1 Vrms sine peaking at 1.414 V. So the README's
"typical" 280 mVpp is −20.1 dBFS peak, and a hot pickup at 500 mVpp is −15.1.
−35 dB is the default, and two things pick it. It gives 5.1 dB of reduction on a hot pickup and 9.7 dB on a humbucker — a compressor doing something the moment you switch it on, which is presumably why you switched it on. And it sits at exactly mid-travel on the new range, so the knob's centre is the sensible setting and both directions mean something.
It is deliberately gentle on a quiet single-coil, 1.4 dB. That is what the 25 dB of travel below it is for.
The other four hold up, with one repair. Ratio's values were always fine —
4.8 is in the useful part of the curve and above 8:1 buys almost nothing — but
the pot was LINEAR(1 20) while the compressor uses 1 − 1/ratio, which
saturates. The first twentieth of the travel covered more than half the effect
and the last 63% covered 0.7 dB. It is EXPONENTIAL(1 20) now: the same
values, the same reach, spread across the knob instead of piled at one end.
Attack costs 5 dB of THD across its whole range while changing the squeeze by more than twice that. Release is right for guitar and marginal an octave below.
Reproducing this
cd Validation
make bench # a stale bench measures a pedal you no longer have
./analyse-compressor.py
Needs ffmpeg for the decode. That decode is deterministic for a given
ffmpeg, and the script prints the sample count, raw peak and scaled RMS
first — if those move and nothing else changed, the decoder did.