Multiband

From the broadcast.audio wiki · Chapter 03 of 9

Part of a series on
Broadcast audio processing

Chapters

Reference

Contents

1The second reason, which matters more
2What a crossover costs
3Where to put the edges
4Two stages per band
5Band linkage, the control that matters
6Listening for it
7See also

A wideband compressor has one gain for the whole signal. Whatever triggers the detector turns everything down — including the things that didn't trigger it.

On a well-made record, that is fine and often desirable. The kick pushes the mix down a little and it comes back up, and the producer chose that. Across a radio playlist it is a defect, and it has a name: spectral ducking. A bass note pulls the vocal down with it. A cymbal crash dips the bass. A truck rumble on a location recording ducks the entire programme for half a second. The listener hears the mix moving underneath itself, and on a station with heavy gain riding it is the single most audible artefact there is.

Splitting the spectrum fixes it in the obvious way: give each region its own detector and its own gain, so the bass note only turns the bass down.

The second reason, which matters more

Preventing ducking is what multiband is usually explained as. It is not actually why stations use it.

Stations use it because it makes different masters sound like the same station. A leveller can make a dull 1973 transfer and a bright 2020 master the same loudness; it cannot make them the same tone. A multiband stage can, because independent per-band gain riding is, functionally, a slow automatic equaliser. If the top band is consistently below its threshold it gets left alone or lifted; if it is consistently over, it gets held down. Over a few seconds the spectrum drifts toward whatever balance the thresholds describe.

This is a very powerful tool and a slightly disreputable one. Used gently, it is the difference between a station and a playlist. Used hard, it takes a record somebody spent a month mixing and reshapes it to a house curve in real time, which is precisely the "re-mastering on the way out" complaint from chapter 1. Both of those are the same control, at different settings. Which is why the amount of linkage between bands is the most consequential number in the whole stage — more on that below.

What a crossover costs

You cannot split a signal into bands for free.

A crossover is a pair (or a set) of filters that divide the spectrum, with the requirement that when you add the outputs back together with no processing in between, you get the input back. Not approximately — audibly exactly, because any error is a permanent EQ curve sitting on your whole output.

Simple Butterworth filters do not do this: split with two 2nd-order Butterworths at the same frequency and the legs are 90° apart at the crossover, so summing gives a 3 dB bump. Linkwitz-Riley solves it by cascading two identical Butterworth sections per leg, giving −6 dB at the crossover in each leg with both legs in phase there, so they sum to unity. Fourth-order Linkwitz-Riley — 24 dB per octave — is the standard choice: steep enough to keep the bands genuinely separate without excessive ringing.

The honest caveat: the sum is flat in magnitude, not in phase. An LR4 split sums to an all-pass — correct level at every frequency, rotated phase around each crossover. Which means a multiband stage that is doing absolutely nothing is still not a wire. On most programme material it is inaudible; on a solo sine sweep or a very transient-heavy signal it is not nothing.

And the bigger caveat: processing the bands and summing is not the same as processing the sum. Once each band has its own gain, the sum is no longer a reconstruction of the input; it is a new signal whose spectral balance moves over time. That is the entire point, but it is also why multiband can sound "wrong" in ways that no individual band sounds wrong.

Two smaller design choices matter. Splitting per channel rather than on a mono sum keeps the stereo information intact inside each band (at the cost of doing twice the filtering). And a cascaded tree — peel off the lowest band, send the remainder to the next split — needs matching all-pass sections on the earlier bands so every band comes out with the same phase history, or the sum stops being flat.

Where to put the edges

Band edges are chosen around how the ear groups sound and where the recurring problems live, not by dividing the spectrum evenly:

Below roughly 100–150 Hz — weight and power. Dominates peak level, barely contributes to intelligibility. This is the band that ducks everything else if you let it.
Roughly 150–400 Hz — body, and also mud and boxiness. The region where a bad room or an old transfer does its damage.
Roughly 400 Hz–1.5 kHz — the core of the midrange. Most of the musical information and most of the perceived loudness. Touch this band the least; it is where processing is most audible.
Roughly 1.5–5 kHz — presence, consonants, intelligibility, and harshness. The ear's most sensitive region. This band is why heavily processed radio sounds forward, and also why it can sound fatiguing.
Above roughly 5 kHz — air, cymbals, sibilance, and hiss. On FM, also the region the pre-emphasis curve is about (chapter 4).

TRANSMIT Tower splits at 120 Hz, 420 Hz, 1.6 kHz and 5.2 kHz — five bands, fourth-order Linkwitz-Riley, split per channel. The edges are fixed and identical on every preset; what changes between presets is what each band does once the signal is there. That is a deliberate choice: crossover frequencies are one of the things that is very easy to make worse and very hard to make better.

Two stages per band

Most broadcast multiband stages run two dynamics stages in series inside each band, because the band has the same two jobs the wideband chain had.

A slow path — a normal compressor with a ratio, attack in tens of milliseconds, release in hundreds — handles the long-term balance. This is the one that does the tonal work described above.

A fast path — effectively an infinite-ratio ceiling with a very short attack, often with a millisecond or two of look-ahead so it can act before the peak rather than after it — stops anything poking out of the band. Because it is band-limited, the distortion it would otherwise cause is confined to that band, and because it is preceded by the slow path it rarely has to do much.

After that, a fixed per-band trim is usually applied, which is simply static EQ: the designer's house tonal curve.

Band linkage, the control that matters

Fully independent bands are a strong effect. Most processors therefore provide linkage (also called coupling): each band's computed gain reduction is pulled toward the average of all the bands' gain reductions by some fraction. At 0 % each band does exactly what its own detector says. At 100 % all bands move together and you have, in effect, a wideband compressor with a spectrally-aware detector.

This one number is the dial between "tonally faithful" and "spectrally aggressive", and it is worth knowing what each end sounds like.

At 0 %, each part of the spectrum fights its own fight. A track with no top end gets its top end brought up; a bass-heavy track gets flattened toward the house balance. Two records that sounded nothing alike come out sounding like siblings. Density is at its maximum, because every band is independently pushed against its own ceiling. The costs are all of the recognisable multiband artefacts: the low band pumping under a kick while nothing else moves; a "hole" where a loud cymbal ducks the 1.6–5 kHz band and takes the vocal's consonants with it; sibilance exaggerated because the top band's gain was up when the "s" arrived; transients losing their snap because the fast path caught them band by band. And the mix balance genuinely moves — a record processed this way is a different record.

At 95 %, the bands move almost as one. You have given up nearly all the automatic-EQ behaviour and kept one thing: the detector is no longer fooled by a single loud region, so a bass note doesn't duck the vocal even though the whole spectrum is still moving together. Tonal balance stays essentially where the mix put it. This is the setting for material where the arrangement is the point — the safe choice, and the one I use for anything acoustic.

Between 30 % and 60 % is where most everyday broadcast settings live: controlled, consistent, still recognisably the record.

TRANSMIT Tower's presets span that whole range, from one preset with the bands completely independent to one at 95 % linkage, and the current figure is shown on the display as a percentage. The reason there is no adjustment for it is the same reason there is no adjustment for the crossovers: linkage is not a value you dial in isolation, it is a value that belongs with a particular set of thresholds, ratios and times.

Listening for it

The artefacts are easier to hear once you know what to listen for. Play something sparse with a strong kick and listen for whether the vocal moves when the kick hits. Play something with a loud ride cymbal and listen to the consonants. Then play a dull old recording and a bright new one back to back, ask whether they now sound like the same band in the same room — and ask whether you actually wanted that.