Parallel compression turns the quiet parts up

The blend that sounded huge

You crush a copy of the drum bus until the meter shows 20 dB of gain reduction and fade it in under the dry drums. The room fills in, the ghost notes appear and the kit sounds twice the size. Somewhere along the way you started treating the blend fader as a percentage: halfway up means half the compression.

Summing a dry path and a compressed path builds a level response of its own, and that response mostly turns the quiet parts up. A 10:1 copy blended in at equal level behaves nothing like a 5:1 compressor.

Why it matters

Two paths, one sum

One path carries the drums untouched. The other carries the same drums through a compressor set hard: fast attack, high ratio, low threshold. The two meet on a bus and add, sample by sample.

On a loud hit, the compressor has pulled the copy far down, so the dry path supplies most of the peak. In the quieter stretch after the hit, the compressor recovers and the copy plays with its full makeup gain, often louder than the dry tail. Added together, the hit keeps most of its height while the decay, the room and the ghost notes come up by several dB.

Listen for two things in the demo as you raise the blend: the space between the hits and the front edge of each hit. The first should change a lot, the second very little.

The science

Adding two levels that move differently

Call the dry level $L_d$ and the copy's level $L_w$, both in decibels. When the two paths carry the same waveform and line up in time, they add in amplitude:

Take a copy compressed at 10:1 from a -30 dB threshold, with no makeup gain and its fader level with the dry track. Below the threshold the two paths are identical, and two identical signals double the amplitude, so the blend is $20\log_{10}2 \approx 6$ dB louder than the dry track. Above the threshold the copy falls behind. With a hard knee, a -20 dB input leaves the compressor at -29 dB (Giannoulis, Massberg and Reiss, 2012), 9 dB under the dry path, and the blend adds only 2.6 dB. A peak at 0 dB leaves it at -27 dB, and the blend adds 0.4 dB.

The "half compression" idea fails this test. A detail at the threshold and a peak at 0 dB are 30 dB apart in the dry track and 24.4 dB apart in the blend. A single 5:1 compressor with the same threshold would squeeze that gap to 6 dB. Right at the threshold, where the blend works hardest, its slope is $(1 + 1/10)/2 = 0.55$, about 1.8:1, and it drifts back toward 1:1 as the input rises. Katz (2002) describes parallel compression as upward compression: the soft passages come up toward the loud ones while the peaks stay put. On real drums, attack and release reshape this static picture (see what attack and release do over time).

The sum only works this way when the paths line up. If the copy arrives late by a time $\tau$, the bus becomes a comb filter (Zölzer, 2011). The paths cancel wherever the delay is an odd number of half periods, at

A 1 ms delay, which is 48 samples at 48 kHz, puts notches at 500 Hz, 1.5 kHz, 2.5 kHz and every 1 kHz above. On quiet passages, where both paths play at the same level, those notches are deep. On loud hits the copy is quieter, so the notches are shallower, and their depth changes as the compressor works.

Most DAWs compensate plugin latency automatically, so a plain software chain usually lines up. The trouble comes where that compensation does not reach: outboard gear patched in without a measured round trip, a plugin that reports the wrong latency, or a low-latency recording mode that bypasses or stops compensating some plugins. An EQ on the copy does a milder version of the same thing: a high-pass shifts the copy's phase around its corner, so the sum can thin out there, as the lesson on filters explains.