Multiband compression and drum punch. Where the front edge goes.

Punch is a crest-factor phenomenon

Ask ten engineers what punch is and you get ten answers, most of them about the low mids. Measure it and it collapses into something simpler: the difference between the short-term peak of a hit and the RMS energy around it. Crest factor. A kick with a 14 dB crest against the surrounding bar reads as punch. The same kick at 6 dB reads as weight, or as loud, or as tiring, but not as punch.

The psychoacoustics behind this are unglamorous. The ear integrates loudness over roughly 100 to 200 ms, but it resolves onset over a far shorter window. The first 5 to 15 ms of a drum hit carries the information the auditory system uses to decide how hard something was struck. Squash that window and the brain reclassifies the event: same spectrum, same average level, less force. This is also why a mix can measure identically on an integrated LUFS readout before and after processing and still feel deflated. Integrated loudness has no opinion about the first 10 ms of anything.

So the working definition for the rest of this page: punch is preserved when the ratio between the onset peak and the sustain behind it survives the chain. Everything below is about the ways that ratio gets spent without anyone deciding to spend it.

Why the broadband compressor gets blamed first

Put a full-band compressor across a dance mix with a 3 dB threshold crossing on every kick and the detector does what it is told: it follows the loudest thing in the program. In most modern productions, that is the low end. A 50 Hz fundamental at −6 dBFS dominates the detector even with a 60 Hz high-pass in the sidechain, because the harmonics of the same event stack across the band.

The audible result is intermodulation of a kind that does not show up in a THD figure. Every kick pulls the entire spectrum down by the gain-reduction amount, then releases it. Hats breathe. A sustained pad pumps at the tempo. If that pumping is the record, fine - a lot of house lives on it. If it is not, you are paying for low-end control with movement in a region that has nothing to do with the low end.

The obvious answer is to split the spectrum so the detector in each band only sees its own material. That is the correct instinct, and it introduces the next three problems.

What the split costs before any gain reduction

A multiband processor is a crossover with dynamics hung off each output. The dynamics get the attention. The crossover does more damage.

Minimum-phase crossovers (Linkwitz-Riley being the usual choice) sum flat in amplitude and are cheap in latency, but each band arrives with its own group delay. A 4th-order LR crossover at 120 Hz rotates phase 360 degrees through the transition region. The kick fundamental and the kick's beater click no longer line up the way they did in the mix. Bypass the processor with all bands set to unity and A/B: on a lot of material you can already hear the kick soften, and a null test against the source will not null. Nothing has been compressed yet.

Linear-phase crossovers fix the alignment and charge you differently. A symmetric FIR has symmetric impulse response, which means energy before the transient as well as after it. Pre-ring. On a 160 Hz crossover with a long filter, that smear sits in the tens of milliseconds and lands in front of the kick, where the ear is at its most sensitive to onset. On sustained material nobody notices. On an exposed 808 or a dry room kick it reads as a soft, slightly hollow attack, and it gets worse the lower the crossover and the steeper the slope. The added latency is a scheduling problem. The pre-ring is a sound problem.

There is no free topology. There is a choice about where to put the error: minimum phase puts it in the alignment between bands, linear phase puts it in time around the transient. On a master bus with drums as the load-bearing element, both choices are consequential, which is why the crossover count and the crossover points matter more than the ratio settings on most multiband processors. Fewer bands, gentler slopes, and crossovers placed away from the fundamental of the loudest percussive element will always out-perform a six-band setup with textbook numbers.

How the band compressor spends the transient

Now the dynamics. A downward compressor in a band is a serial device: the whole band passes through the gain element. Whatever it does to the sustain, it does to the attack first, because the attack arrives first.

Three mechanisms eat punch here, and they compound:

• —Attack time. A 1 ms attack catches the onset itself, which is exactly the part you wanted to keep. A 30 ms attack lets the transient through and clamps the body, which raises crest factor and can read as more punch - until the release timing starts modulating the tail.
• —Detector type. A peak detector on a low band follows the waveform of a 50 Hz sine, which completes a cycle in 20 ms. Anything faster than that is ripple, and ripple in a low-band gain signal is distortion at the fundamental. RMS or a longer-window detector avoids it and gives up onset accuracy in return.
• —Release interaction with tempo. Release times in the 80 to 300 ms range interact with the grid. At 128 BPM a sixteenth is 117 ms. Set the release near that and the gain signal locks to the groove, either reinforcing it or fighting it. This is the single most under-measured parameter in mastering dynamics.

Note what all three have in common: the gain signal is a function of the band's own recent history, and the whole band is subject to it. That is the structural cost of serial processing. You cannot control the sustain without touching the onset, because there is one signal path and one gain element.

Why parallel band processing behaves differently

Parallel topology changes the arithmetic. Instead of routing a band through a gain element and passing the result on, you keep the source band intact and sum a processed copy alongside it. The dry path is a straight wire. The transient in the dry path is untouched by definition.

What the processed copy contributes is sustain energy: a heavily compressed, slower-attack version of the same band that fills in behind the hit. Sum them and the peak stays where it was while the RMS rises. The crest factor drops, but it drops from below - by adding body, not by removing edge. Measured on a kick, you typically see integrated level up 1.5 to 3 dB with the sample peak within a few tenths of a dB of the source. That is the entire trick, and it is why parallel band gain survives a limiter downstream in a way serial multiband does not.

There is a cost, and it is phase again. The processed copy must be sample-aligned with the dry path or the sum comb-filters. In a band-split parallel design that means the compressed path cannot introduce any latency the dry path does not also carry, and any non-linearity in the processed path (saturation, soft knee, harmonic generation) will move the phase relationship at frequencies it colours. Designs that get this wrong sound thick and slightly detached, as if the low end were arriving from a different room. Designs that get it right null cleanly with the wet path at zero.

The other reason parallel wins on the mix bus is failure mode. When a serial multiband is pushed too hard, the mix collapses. When a parallel band is pushed too hard, it gets fat and eventually muddy, which is a fault you hear immediately and back off from. Given a choice, prefer the processor whose overdriven state is obvious.

What to look at, and what to ignore

Most of the argument above is settled with four numbers and one discipline. The discipline first: every comparison is loudness matched, and matched on integrated LUFS, not on peak and not by ear with the fader. An unmatched A/B of a multiband processor is a test of whether louder sounds better, and it always does, for about 90 seconds.

Symptom

Kick loses front edge, mix gets louder and smaller

Root cause

Crest factor collapse in the 40-120 Hz band

Measure

Per-band crest factor, PLR, short-term vs integrated LUFS

Crossover cost

Minimum phase: group delay. Linear phase: pre-ring

Fix

Parallel band gain, dry transient path kept intact

Verify

Loudness-matched A/B, never level-matched by fader feel

Per-band crest factor is the one that answers the punch question directly. Measure the 40-120 Hz band before and after. If crest dropped and the master got louder, you bought loudness with the transient. If crest held and integrated level rose, the processing added sustain rather than removing onset.

PLR (peak to loudness ratio) is the whole-file version of the same question and is the number worth quoting to a client, because it survives platform normalisation. A master at −9 LUFS-I with a PLR of 8.4 has meaningfully more headroom left in the transients than the same track at −9 with a PLR of 5.5, and on a normalising platform they will be played back at the same loudness. One will sound like a record and one will sound like a limiter.

Short-term LUFS variance across the arrangement tells you whether the processing flattened the structure. If the drop and the breakdown converge to within 1 dB short-term, the arrangement has been processed out of the record, whatever the spectrum says.

Inter-sample true peak is the housekeeping number. Multiband processing raises band energy near crossover points and a 0.0 dBFS sample-peak master can easily be +1.2 dBTP after codec conversion. Leave true peak at −1.0 dBTP for streaming delivery unless there is a specific reason not to.

What to ignore: real-time spectrum analysers as an arbiter of punch. A spectrum plot integrates over a window and shows you none of the time-domain behaviour that this entire page is about. Two masters with identical average spectra can differ by 6 dB of crest factor in the kick band.

A working order of operations

For a mix that arrives with the low end broadly right but soft in the front edge, this is the order that costs the least:

• —Fix static balance first. If a 1.5 dB shelf at 60 Hz solves it, no dynamics processor is needed and none should be inserted. Dynamics applied to a static tonal problem is the most common self-inflicted wound in mastering.
• —Choose crossover points against the material, not the preset. Find the kick fundamental, then place the nearest crossover at least half an octave away from it. A crossover sitting on the fundamental of the loudest element in the record is the worst available choice.
• —Set the sustain path, not the peak path. Slower attack, release chosen against the tempo grid, and enough ratio that the compressed copy is genuinely dense. It is a texture generator, not a safety device.
• —Blend by measurement, then by ear. Bring the parallel band up until per-band crest starts to move, then back off a step and listen at matched loudness.
• —Check mono and check the sub. Parallel low-band energy that only exists in the sides will disappear on a club sum and on most phones. Kick and bass relationship is verified in mono before anything else is signed off.
• —Limit last, and check what the limiter had to do. If the limiter's gain reduction went up after the multiband stage, the multiband stage raised peaks rather than sustain. That is a topology error, not a settings error.

Where ThuMP sits in this

Everything above is topology and measurement, and it is implementable with whatever is already on your system. ThuMP exists because doing it this way by hand across a lot of records turned into a fixed procedure, and a fixed procedure is a specification.

ThuMP is a master-bus processor built around parallel band gain rather than serial band compression. The dry path stays intact, the processed copies are sample-aligned, and the controls are the ones that matter in the procedure above: how much sustain each band contributes, how fast it arrives, and how it relates to the tempo of the material. Loudness is the outcome of that, not the target.

LEARN is the part that sets a starting point. It is a Bayesian optimiser: it measures the material, searches the parameter space against measured objectives, and proposes settings. It is an algorithm, not a trained model, and it is not AI. It has been run against real music and signed off in this room, on the hybrid analog chain, by an engineer who ships records. That is the honest ceiling of what it does. It gets you to a defensible starting point in a few seconds. The last 2 dB is still a judgement call, and it still belongs to you.

MuMP handles the same thinking upstream on the mix bus, where glue matters more than weight. LuMP handles delivery, where the only question left is whether the file survives the platform. The measurement discipline is shared across all three, because it is the only part that is not a matter of taste.

Six sentences

• —Punch is crest factor in the drum band, not a frequency you can boost.
• —Full-band compression lets the low end modulate the whole spectrum.
• —The crossover costs you phase alignment or pre-ring before any gain reduction happens.
• —Serial band compression cannot control sustain without touching onset, because there is one gain element.
• —Parallel band gain raises RMS while leaving the peak where it was, which is what survives the limiter.
• —Loudness is the outcome. The objective is that the record still moves.