Limiter release reaches into the groove

The hats that started breathing

The mix bounces clean. You put a limiter on the master, push it until the meter reads where you want, and the groove changes. The hats and the pad now swell up between kicks and duck on every downbeat, like a sidechain nobody set up. You lengthen the release to stop it, and the master gets quieter. You shorten it, and the bass picks up a furry edge.

The limiter did both of those things through one control. Its release decides what the rest of the mix is doing between the hits that triggered it.

Why it matters

The limiter turns everything down at once

A master limiter sees one signal, the whole mix. When a kick crosses the ceiling, the limiter cannot turn down the kick alone. It turns down the bass, the pad, the vocal and the hats that are playing at that moment, then lets them back up over the release time. Each kick leaves a dip in everything else, and the release sets its shape.

On a drum loop in a mix session, release decides how the next hit lands. On a master the more audible effect is often on the parts that do not trigger the limiter at all: the sustained bass under the kick and the off-beat hats that fall inside the dip.

The science

The recovery curve against the beat

A digital limiter measures the peak level, computes how much gain reduction is needed to keep it under the ceiling, and smooths that gain over time (Zölzer, 2011). Many limiters add a short look-ahead delay so the gain is already down when the peak arrives. Recovery is commonly modelled as a one-pole smoother on the gain reduction in decibels (Giannoulis, Massberg and Reiss, 2012). Once the kick has passed, the gain reduction left after a time $t$ is:

Here $GR_0$ is the reduction on the kick and $\tau$ is the release time constant. The time to compare it with is the gap between beats:

At 120 BPM a beat lasts 500 ms, so an off-beat hat lands 250 ms after the kick. With 4 dB on the kick, a 50 ms release has almost fully recovered by then. A 250 ms release still holds the hat $4 \times e^{-1} \approx 1.5$ dB down, and the bed rises by about 3.5 dB across each beat: that rise is the pumping you hear. A 1 s release holds everything about 3 dB down nearly all the time. That is clean and steady, but you need more drive to reach the same loudness.

MATLAB's Audio Toolbox, for example, defines attack and release as the time the gain takes to move from 10% to 90% of its final value, which for a one-pole smoother is $\tau \ln 9 \approx 2.2\tau$. Two plugins can both read 250 ms and recover at different speeds, so set release by ear against the tempo, not by number.

Go the other way and the release becomes faster than the bass. A 50 Hz cycle lasts 20 ms, with a crest every 10 ms. If the gain recovers within a few milliseconds, it rises and falls with every crest, and the limiter starts reshaping the wave instead of riding its level. In my simulation of a 50 Hz sine driven 6 dB into the ceiling, a 1 ms release produced a third harmonic about 15 dB under the fundamental, against about 39 dB under with a 100 ms release. The fast version also came out about 1.3 dB louder at the same peak level, because it was behaving like a clipper.

Some limiters split the job between two stages: a fast one that catches short peaks and a slower release stage that follows the average level. FabFilter's documentation for its Pro-L 2 limiter describes that design and its trade-off: short attack and long release settings are "safer and cleaner" but "can also cause pumping and reduce clarity", while long attack and short release can raise apparent loudness "at the expense of possible distortion". Program-dependent and auto modes make that trade-off for you, so their choice still needs checking by ear.

Listen to the space between the hits as you change the release, with the loudness held equal.