Compressor signal paths

Switch between feed-forward, feedback, sidechain, parallel, multiband and stereo link in a block diagram and compare them.

Explanation Tasks

What is happening?

Same building blocks, different paths

Every compressor consists of the same building blocks: a gain stage (VCA, voltage-controlled amplifier) in the audio path that changes the gain, and a control path of detector, transfer curve and attack/release that works out how much to reduce. Here the question is where the control path takes its signal from and how the paths are connected – this often shapes the behaviour more than the control settings.

The “Block diagram (click a block → its controls)” shows the selected signal path. Audio paths are solid, control paths (level and gain in dB) dashed, signals that only feed a detector are in the input colour. Hover over a block to see an explanation below it; clicking highlights its controls. Switching the “Topology” selects a suitable test signal.

Feed-forward and feedback

In a feed-forward compressor the detector measures the input and calculates the reduction forwards. This allows exact transfer curves, ∞ : 1 and lookahead. In a feedback compressor the detector measures the output, i.e. the already reduced signal: the control loop sees its own result. Many classic opto and tube compressors work this way.

If the input lies i dB and the output o dB above the threshold and the controller reduces by k · o dB (k is the control strength: reduction in dB per dB that the output lies above the threshold), then o = i − k · o, so o = i / (1 + k). To reach the ratio R (the output rises by 1 dB when the input rises by R dB, i.e. o = i / R), the feedback controller must therefore reduce by (R − 1) dB per dB of overshoot. Example 4 : 1 with the sine burst: input 14 dB above the threshold, output 3.5 dB, reduction 3 · 3.5 = 10.5 dB – statically the same result as with the feed-forward compressor. ∞ : 1 would need an infinite slope, and lookahead is impossible because the output does not exist yet; the model therefore allows at most 20 : 1 (readout “Effective ratio”).

The real difference lies in the timing: because the time constants sit inside a loop, even a small change in output level triggers a large change in reduction. Try it: topology “Feedback”, section “Attack”, ratio 4 : 1, knee 0 dB, attack 10 ms, release 150 ms. The dashed curve “GR feed-forward” (GR = gain reduction, the reduction in dB) shows a feed-forward compressor with the same detector. It needs about 38 ms to reach 90 % of the reduction, feedback only about 12 ms; after the loud part the reduction falls below 1 dB after about 225 ms with feed-forward, but after only about 56 ms with feedback. At 10 : 1 feedback takes about 7 ms and 23 ms.

Sidechain: what the detector hears

In the sidechain the detector receives a different or a filtered signal from the one being processed. A common misconception: a sidechain filter changes the sound of the output. It does not – it only changes what the detector responds to.

Try it: topology “Sidechain” with “Kick + pad” (threshold −24 dB, 4 : 1, attack 5 ms, release 200 ms). The kick reaches about −3 dBFS, the pad only about −14 dBFS; so the kick determines the detector level. With every hit the pad is pulled down as well and comes back up afterwards – the readout “GR fluctuation (0.5–1.5 s)” shows about 9 dB, the pad pumps. Now set “Sidechain filter” to “High-pass” at 150 Hz: the fluctuation drops to just under 2 dB, the pad is compressed gently and evenly, and the kick is still fully present in the output.

The detail view “Sidechain filter in the detector path” shows why: the curve “What the detector hears” is the filter’s frequency response, and the ranges of kick and pad are marked below it. The high-pass (two filter stages, 24 dB/octave) is 6 dB down at its filter frequency, about 22 dB at 80 Hz where the kick starts – but only 1.7 dB at the lowest pad note (220 Hz). A “Band-pass”, by contrast, only lets a range around the filter frequency through; placed at 5–8 kHz, the compressor reacts only to sibilants and becomes a de-esser.

Ducking with an external key

Under “Key signal (drives the detector)” you can give the detector an entirely different signal: “Kick” or “Speech”. Only the input is still processed and output; the key signal appears in the time history as an additional trace “Key”. This creates ducking: the music makes way for the presenter, the bass for the kick. In electronic music, rhythmic pumping via a kick sidechain is even a deliberate effect.

Try it: “Pad + bass” as the signal, key “Speech”, threshold −30 dB, 4 : 1, attack 5 ms. As soon as someone speaks, the music drops by up to about 14 dB. Compare release 50 ms, 150 ms and 400 ms: with 50 ms the music comes almost fully back between syllables (fluctuation about 14 dB) – which sounds restless. With 400 ms it stays at least about 5 dB down while speech continues and only returns in longer pauses (fluctuation about 9.5 dB). For presenting over music, long release times are therefore preferred. Set the key to “Kick” and the music pumps in time with the kick, although the kick does not appear in the output at all.

Parallel compression

In parallel compression (“New York compression”) an unprocessed path (dry) and a heavily compressed path (wet) are added. The result is upward compression: quiet parts are lifted, loud ones stay almost unchanged, and the transients survive in the dry path. A common misconception is that this is the same as a lower ratio: a normal transfer curve pulls loud levels down, the parallel mix lifts quiet ones.

Worked example with the settings applied when switching to “Parallel” (threshold −35 dB, 10 : 1, make-up 10 dB, dry 0 dB, wet −6 dB): an input of −40 dB lies below the threshold; the wet path delivers it at −40 + 10 − 6 = −36 dB. In-phase signals add linearly: 10^(−40/20) + 10^(−36/20) gives −31.8 dB, i.e. +8.2 dB. An input of −6 dB is reduced by 26.1 dB in the wet path and then lies 22 dB below the dry signal; the sum is only 0.7 dB louder. At the top of the detail view “Parallel mix: transfer curve and comb filter” you see this as the curve “Dry + wet”. With the “Drum groove” the peak level therefore rises only from −2.6 to about −2.1 dBFS.

This only works if both paths arrive at the same time. If the compressor’s latency (e.g. from lookahead or oversampling) is not compensated, summing creates a comb filter with cancellations at odd multiples of 1 / (2 · latency). Try it: switch “Delay compensation” off. With 3 ms latency the “First cancellation” lies at 1 / (2 · 3 ms) ≈ 167 Hz, the next ones at 500 Hz and 833 Hz – right in the fundamental range of kick and snare. How deep the notches get depends on the level ratio of the paths; they can cancel completely only at equal levels.

Multiband: one compressor per band

A multiband compressor splits the signal into bands with a crossover and gives each band a compressor of its own. Here there are three; their boundaries are set by “Crossover low / mid” and “Crossover mid / high”. Each band has its own threshold and ratio controls; knee, attack, release and detector apply to all bands, and make-up gain acts after the sum.

Try it: topology “Multiband” with “Kick + pad” (crossovers 150 Hz and 2 kHz, all thresholds −24 dB, 4 : 1). “GR low” fluctuates by just over 12 dB in time with the kick, “GR mid” by less than 2 dB: the kick now only pulls down the lows, the pad stays almost untouched. The detail view “Bands with current gain reduction” shows the bands without compression dashed, the bands with the reduction at the cursor solid and the “Sum” bold. With “Listen to” you solo a single band.

The Linkwitz-Riley crossover (24 dB/octave) sums to a flat magnitude response but shifts the phase (all-pass): uncompressed the signal sounds almost the same, the peak values change slightly. As soon as the bands are working, the multiband compressor acts like a level-dependent EQ. But it is not simply “a better EQ”: it only acts when a band exceeds its threshold – in mastering, for example, on a bass that only booms in loud passages – and so it can shift the balance between quiet and loud passages.

Stereo link

With “Detectors: Unlinked” each channel measures and is controlled on its own. This shifts the stereo image: whatever sits in the centre is equally loud on both channels; if only one channel gets quieter, it drifts to the opposite side.

Try it: topology “Stereo link” (threshold −20 dB, 4 : 1, attack 5 ms). The test signal contains a snare on the left only and a pad in the centre. Unlinked, only the left channel is reduced with every snare hit (by up to about 10 dB, the right by just over 2 dB from the pad itself). The pad is then quieter on the left than on the right and drifts up to about 8 dB to the right – the detail view “Stereo image of the centred pad” shows it jumping about on the L … R (left … right) stage, “Largest shift to the right” the maximum. Switch to “Linked”: now the louder channel (block “Maximum”) sets one common reduction for both. The pad stays in the centre; in exchange, the right side also gets quieter with every hit.

On the mix bus and on stereo groups the detectors are therefore almost always linked. Unlinked operation suits two independent signals, such as two microphones in dual-mono mode.

Limits of the model

All variants run at a 16 kHz sample rate and are simplified; real devices differ in detector, curve and colouration. So that only the position of the detector makes the difference when comparing feed-forward and feedback, both use the same envelope detector there (attack when the measured level rises, release when it falls); the feedback detector measures the output before make-up gain, delayed by one sample. The topology “Feed-forward”, on the other hand, uses the control loop from “Compressor and limiter” and behaves slightly differently in time.

Lookahead is only available with feed-forward; partial linking or a mid/side mode, as on many real devices, are not available at all. The latency in the parallel path is a pure delay; the multiband crossover consists of IIR filters and is not linear-phase – linear-phase crossovers avoid the phase shift but cause latency and pre-ringing. Peak values are measured only at the samples (no true peak).

Tasks

Practice questions

Questions from studio, stage and mastering – to think about, discuss and check with the module.

1. A speaker's sibilants are harsh. What is the difference between a de-esser with a band-pass in the sidechain and a multiband compressor with a band at 5–8 kHz?

Show answer

With a sidechain de-esser only the sibilant triggers the reduction, but the whole signal is reduced – set heavily, the voice also gets duller and quieter on every S (“lisping”). The multiband compressor reduces only the upper band and leaves fundamental and formants alone; in return the crossover and its phase response act on the signal. Many de-essers therefore offer both modes (wideband / split band). A short attack and moderate reduction (3–6 dB) keep it natural.
2. In mastering the bass should be tamed in loud passages. Why is a multiband compressor not simply “a better EQ” here, and why does it sound slightly different even at ratio 1 : 1?

Show answer

An EQ always acts the same way, the multiband compressor only when the band exceeds the threshold – it is a level-dependent EQ. That is the point, but it can shift the balance between quiet and loud passages. Even at 1 : 1 the signal passes the crossover: the Linkwitz-Riley bands sum to a flat magnitude, but with phase rotation (all-pass) around the crossover frequencies – peak levels and transients change slightly. Linear-phase crossovers avoid this but add latency and pre-ringing. So: place the crossovers deliberately, only actively process the bands that need it and always compare with bypass at equal loudness.
3. A bus compressor on the stereo sum: should the link be on? And what changes in mid/side mode?

Show answer

On the sum almost always linked: otherwise vocals, bass and kick in the centre drift back and forth whenever one side is loud (hi-hat left, guitar right). Unlinked operation suits two independent signals, such as dual mono for two microphones. Many devices offer intermediate settings (partial link) or a sidechain mix. In M/S mode, mid (L + R) and side (L − R) are compressed separately: reducing the mid makes the mix wider, reducing the side narrower – so the stereo image changes here too, only deliberately.

Tasks

Predict first, then try it out, then open the solution. “Load starting point” sets the matching initial values.

1. Feedback acts faster

At 0.3 s the sine jumps from −30 to −6 dBFS; ratio 4 : 1, attack 10 ms, release 150 ms. Compare the gain reduction curve of the feedback compressor with the dashed feed-forward curve (same detector). How long does each take to reach 90 % of the reduction? Then switch to the “Release” section and finally set the ratio to 10 : 1.

Load starting point

Show solution

Both settle at the same output level (about −15.8 dBFS), so the transfer curve is the same. But feed-forward needs about 38 ms to reach 90 % of the reduction, feedback only about 12 ms; after the loud part the reduction drops below 1 dB after about 225 ms with feed-forward and after about 56 ms with feedback. At 10 : 1 it becomes about 7 ms and 23 ms: because the controller measures its own result, attack and release act faster the higher the ratio.
2. Pumping and sidechain high-pass

Kick and pad run through one compressor. Watch the gain reduction and the “GR fluctuation” readout: the pad gets quieter with every kick. Then switch on the high-pass in the sidechain (150 Hz).

Load starting point

Show solution

Without the filter the reduction fluctuates by about 9 dB with the kick – the pad pumps. With the high-pass the detector barely hears the kick, the fluctuation drops to just under 2 dB; the pad is compressed lightly and evenly. The kick is still fully present in the output, because only the control signal is filtered.
3. Ducking with an external key

The music (pad + bass) should get quieter whenever someone speaks; the speech signal drives the detector as the key. Compare release 50 ms, 150 ms and 400 ms.

Load starting point

Show solution

The music is reduced by up to about 14 dB. With 50 ms release it comes back almost fully between syllables (fluctuation about 14 dB) – which sounds restless. With 400 ms it stays down by at least about 5 dB while speaking and only returns in longer pauses (fluctuation about 9.5 dB). For voice-over music, longer release times are therefore usually chosen.
4. Parallel compression and latency

The drum groove is compressed in parallel (10 : 1, make-up 10 dB, wet −6 dB). Read from the curve how much an input of −40 dB and of −6 dB is lifted, and compare input and output peak level. Then switch off delay compensation.

Load starting point

Show solution

An input of −40 dB comes out at about −32 dB (+8 dB), −6 dB only at about −5.3 dB (+0.7 dB): quiet parts are lifted strongly, loud ones hardly – the peak level only rises from −2.6 to about −2.1 dBFS. Without delay compensation the wet branch arrives 3 ms late; the sum gets a comb filter with cancellations at about 167 Hz, 500 Hz, 833 Hz … – right in the fundamental range of kick and snare.
5. Stereo link

A snare plays on the left, the pad sits in the centre. With unlinked detectors, watch the position of the pad in the detail view. Then switch to linked detectors.

Load starting point

Show solution

Unlinked, only the left channel is reduced with each snare hit (up to about 10 dB, the right side only just over 2 dB from the pad itself). The pad is then quieter on the left than on the right and drifts up to about 8 dB to the right – the stereo image wobbles. Linked, both channels are reduced equally and the pad stays in the centre; in return the right side also gets quieter with each hit.

Settings

Sound

Level calibration (dB SPL)

MIDI keyboard

Compressor signal paths