torvalds-GuitarPedal/Documentation/effects/reverb.md

Reverb [REVERB]

Freeverb — Schroeder-Moorer, by Jezar at Dreampoint, public domain. Eight parallel feedback comb filters fed by the input, summed into four series allpasses. Two controls: Room (0.70..0.98) is the feedback gain shared by all eight combs, and Damp (0.1..0.5) is a one-pole lowpass inside each comb's feedback loop. Defaults are 0.88 and 0.25, and it returns wet only, at an 18% default mix.

Each comb's read pointer is swept ±6 samples by one of four slow LFOs, which is the Lexicon trick for stopping a long tail ringing on a fixed set of frequencies.

Room is the decay time, and it is not linear in the knob

T60 is measured the standard way: Schroeder backward integration of the tail after a gated noise burst, fitted from −5 dB to −25 dB and tripled.

%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2'}}}}%%
xychart-beta
    title "Decay time against the Room control"
    x-axis "Room, comb feedback gain" [0.7, 0.76, 0.81, 0.87, 0.92, 0.98]
    y-axis "T60, seconds" 0 --> 10
    line [0.52, 0.65, 0.89, 1.32, 2.4, 9.08]

Half a second to nine seconds, so the whole travel does something — but look at the shape. The knob is linear in the feedback gain, and decay time goes as roughly 1/(1−g), so it runs away at the top: the first 80% of the pot covers 0.52 to 2.4 seconds, and the last 20% covers 2.4 to 9.1.

That is not merely a steep curve, it is a curve whose steep end is where the interesting settings live. It is the same shape of defect the compressor's Ratio had — a control whose useful range is piled at one end of its travel — and it is worth knowing about before reaching for the knob, because the top two centimetres are where a room turns into a cathedral.

The default 0.88 gives 1.32 s, which is a plausible medium room and sits at mid-travel where the control is still well behaved.

Damp tilts the decay; Room scales it

Damp is a lowpass in the feedback path, so it does not shorten the tail evenly — each pass round the loop takes a little more top off, and the highs run out first. That is what a real room does, because air and soft furnishings both absorb treble faster than bass.

%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00, #009e73'}}}}%%
xychart-beta
    title "T60 by octave. Damp 0.10 blue, 0.30 orange, 0.50 green"
    x-axis "Hz" [125, 250, 500, 1000, 2000, 4000]
    y-axis "T60, seconds" 0 --> 2
    line [1.7, 1.85, 1.77, 1.69, 1.47, 1.13]
    line [1.7, 1.84, 1.71, 1.6, 1.24, 0.78]
    line [1.69, 1.81, 1.59, 1.47, 0.91, 0.48]

The three lines are on top of each other at 125 Hz — 1.70, 1.70, 1.69 — and a factor of 2.4 apart at 4 kHz. Damp does nothing whatever to the bottom of the spectrum, by construction, and that is the correct behaviour rather than a limitation.

Room, by contrast, scales the whole thing:

%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2, #d55e00, #009e73'}}}}%%
xychart-beta
    title "T60 by octave. Room 0.70 blue, 0.87 orange, 0.98 green"
    x-axis "Hz" [125, 250, 500, 1000, 2000, 4000]
    y-axis "T60, seconds" 0 --> 11
    line [0.95, 0.59, 0.59, 0.58, 0.56, 0.46]
    line [1.58, 1.68, 1.5, 1.51, 1.21, 0.82]
    line [10.82, 10.12, 9.6, 7.68, 4.62, 1.79]

Note that the green line still slopes — at Room 0.98 the tail is 10.8 s at 125 Hz and 1.8 s at 4 kHz, a six-to-one tilt, because the damping compounds over far more trips round the loop. So the two controls are not independent: a longer room is also a darker one, at the same Damp setting.

The modulation, and why the tail wanders

Eight fixed combs ring on a fixed set of frequencies, and a note that lands on one of them sits there and hums. Sweeping each read pointer ±6 samples walks those resonances instead — ±0.5% on a comb of 1215 samples — so nothing has a fixed frequency to sit on.

You can see it directly. A steady 440 Hz tone, 100% wet, half-second windows — the first 5.5 seconds of a 85-second capture:

%%{init: {'themeVariables': {'xyChart': {'plotColorPalette': '#0072b2'}}}}%%
xychart-beta
    title "Wet level of a steady tone, wandering as the combs detune"
    x-axis "seconds" [0.0, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0, 4.5, 5.0]
    y-axis "dBFS" -28 --> -24
    line [-25.08, -26.57, -25.92, -24.96, -26.08, -25.12, -25.73, -26.11, -26.86, -25.24, -25.96]

The input never changes and the output wanders. That wander is the effect working: it is the comb peaks sliding under a fixed tone.

How far it wanders depends entirely on how long you watch, and that is worth stating because the first attempt at this got it wrong. The excerpt above covers 1.9 dB. Over the full 85 seconds it is 5.9 dB peak to peak, with a standard deviation of 1.08 dB and a p5..p95 spread of 3.4 dB. The slowest of the four LFOs has a period of 4.8 seconds and they are spaced so as not to share one, so a short capture does not sample the range — it samples whichever corner of it the recording happened to start in. An eight-second capture reported 2.3 dB, which is less than half the truth.

The chart is an excerpt rather than the whole run for a reason that is about drawing rather than about reverb: eleven points is what the axis can label, and eleven points spread over 85 seconds would put the window well above every rate being looked at and average the wobble into a flat line.

The four LFOs run at 0.21, 0.31, 0.46 and 0.67 Hz — spaced about 3:2 so they do not beat into a common period — and they are phase accumulators, which matters more than it sounds. They used to be quadrature phasors rotated by a fixed step, and nothing renormalised them: three of the four spiralled outwards and one decayed to a tenth of its amplitude in ten minutes, so the reverb slowly became the unmodulated version of itself and, left on overnight, walked its read pointer out of the comb buffer entirely.

The modulated read is interpolated. Truncating it to a whole sample makes the pointer jump, and a jump is a step discontinuity sprayed into the tail continuously by eight combs at four rates. Measured by band-limiting the input to 1 kHz and looking above 4 kHz, where a reverb that is time-invariant apart from a sub-hertz modulation cannot legitimately put anything:

  input above 4 kHz    -78.3 dB of its total
  output above 4 kHz   -82.2 dB of its total

The output has less up there than the input did. It manufactures nothing. Before the read was interpolated it manufactured 34.5 dB above what went in.

What it costs

This is the expensive effect, and unlike everything else on this page the number comes off the pedal rather than the bench — Validation/measure-load.py, which reads the load meter over MIDI telemetry. Steps out of 127 of the sample period, averaged over four boots:

	reverb routed	share of the sample period
empty chain	—	9.4 %
reverb, as it is now	28.42 steps	22.4 %
before the LFO and interpolation fixes	26.15	20.6 %
with those fixes but a per-frame LFO	33.54	26.4 %

So one reverb is a fifth of the audio budget, which makes it comfortably the most expensive effect here and the one to think about before stacking things behind it.

Those four figures were taken with the load meter as it was, timing the audio loop's idle spin with a 1 MHz timer — one tick of which is 4.8% of a sample period, so each of them is a rounding of that. The meter counts cpu cycles now, and re-measured it puts the reverb at 22.1%, which is inside the old instrument's resolution and does not move anything said here. The before-and-after rows have not been re-measured, because doing so means reflashing three superseded builds to correct figures by a few tenths.

The middle row is the honest cost of correctness: fixing the drifting LFOs and interpolating the comb reads added 28.3%. Running the LFOs at control rate — exact once every 32 frames, a straight line between — gave most of that back, and the effect now costs 8.7% more than the broken version rather than 28.3%. The remaining work is the comb read itself, which is irreducibly per-frame because it reads audio.

Reproducing this

cd Validation
make bench          # a stale bench measures a pedal you no longer have
./analyse-reverb.py

The cpu figures are not from that and will not be reproduced by it. They need the pedal:

./measure-load.py -b 4

which sets routing live over SysEx, saves nothing, and reloads the scene when it finishes. Read the note at the top of it before believing a single reading: the empty-chain baseline wanders by a couple of steps between boots while the routed reading holds still, so the number to compare between firmwares is the routed absolute, averaged over several boots, and never a one-shot difference.