808 and kick: what 1016 files taught us about reading the low end

Before a tool can help with your 808, it has to know there is one

Almost every article about an 808 clashing with a kick starts from the same place: it assumes you have one. That is fine for an article you chose to open. It is not fine for a tool that has to decide, on its own, whether the low end in front of it is an instrument carrying notes or a drum carrying a transient, because the advice for those two is different and the wrong one is worse than silence.

So the first thing we measured was not how to fix a clash. It was the gate: what shape of measurement means this material's low end is the instrument.

Key takeaways

• We tested four candidate gates against the same 1016 files, using the real production analysis rather than a copy of it. Zero files failed to process.
• The gate we shipped is the plainest one: the loudest band in the table is the sub band, 20 to 60 Hz. No threshold, no calibration, no tuned number.
• It opens on 88 of 1016 files, which is 9 percent. Of those, 10 percent have a filename that mentions the low end, against a corpus base rate of 2.4 percent, so roughly four times the base.
• The clearest evidence is six files with 808 in the name. The gate opened on four and stayed shut on two, and the two were a clap and a snare roll.
• The band table cannot tell you what is making the low end. An 808, a sine sub, a bass guitar and a long tailed kick look the same in it.

Sub and bass are two different jobs

The measurement splits the spectrum into six bands. Two of them matter here.

Treating those as one thing sends every decision to the wrong place. A mix that is all sub and no bass reads enormous on headphones and empty everywhere else. A mix that is all bass and no sub reads small on a club system. The words low end cover both and hide the difference.

The run

The corpus is a working sample library, 1016 wav files, measured on 15 August

1. The chain was deliberately split across two sides: the application produced the raw analysis payloads with the same code that ships, and the backend scored them with the same functions that run in production. Neither side marked its own homework. Zero of the 1016 files dropped out.

Four candidates were scored against that one run.

Hit rate here means: of the files where the gate opened, how many have a filename that mentions the low end. Filenames are a weak label. They are sparse, they are inconsistent, and plenty of low end material is named something else entirely. So the absolute values are not the finding. The comparison is.

The base rate across the whole corpus is 2.4 percent. The gate we shipped runs at 10 percent, which is roughly four times that. The widest gate, the one that opens on 38 percent of everything, runs at 4 percent, which is barely above the corpus itself. A gate that opens on more than a third of all material is not detecting anything.

The six files that make it readable

Six files in the corpus carry 808 in the name. The gate looked at the measurement, never at the name.

The two it stayed shut on are a clap and a snare roll. In both, the 808 in the name refers to the drum machine, and neither has anything to do with the low end. A gate that read filenames would have got both wrong. This one reads the material.

Why there is no number in the gate

The shipped gate is a ranking, not a threshold. It asks which band is loudest and checks whether the answer is sub. There is no dB value in it anywhere.

That was the point. A gate with a threshold in it creates a calibration debt on the day it is written, and that debt goes stale quietly: the number keeps working, nobody rechecks it, and two years later it is drawing a line somewhere nobody chose. A ranking has no such debt.

It also has a property we tested rather than assumed. Shift the entire band table down by 20 dB and the decision does not change, because every band moved together and the ordering is intact. A quiet loop and a loud one get the same reading. This is the same failure mode we wrote about when an analyzer calls a mix clean and your ears do not: absolute level thresholds report on how loud something was played, not on how it is balanced.

What the measurement refuses to say

Two refusals, and both are written into the tool rather than left to judgement.

It does not name the instrument. An 808, a sine sub, a bass guitar and a kick with a long tail all produce the same shape in a band table. Unless you have said what it is, the tool says the low end. Naming an instrument it cannot actually identify is the kind of confident wrong thing that costs every answer after it.

It does not name a frequency. The measurement produces bands, not a spectrum. Cut 47 Hz is a sentence that claims a resolution the reading does not have. Under about 40, or somewhere around 200, is what the data supports.

What we do with it once the gate is open

The gate decides whether to talk. What gets said is background knowledge, not a reading of your track, and it is worth separating those two clearly.

Two sources in one octave is a decision, not a boost. A kick and a sustained low note occupy the same range and mask each other. Boosting both makes it worse, because they are competing for the same energy rather than lacking it. The question is who owns the fundamental. Usually one is short and one is long, or one steps out of the way while the other sounds. Any of those work. Two boosts do not.

Length is a level control. In a busy arrangement, how long each low note rings often matters more than how loud it is. Overlapping tails add up, and the loudest moments are the ones where the previous note has not finished. Shortening the decay clears more room than pulling the fader, and it does it without making the part smaller.

Turning it up is the wrong fix for a phone. Phone speakers, laptop speakers and most earbuds put out almost nothing below about 100 Hz. The fundamental is simply not reproduced. People still hear the pitch, because the ear reconstructs it from the upper harmonics. So raising the level does not help: the speaker still cannot play that band, and the extra energy now eats headroom on every system that can. The move that works is harmonic, meaning gentle saturation on the low part or a quieter layer an octave up following the same notes, so a small speaker has something to play. Judge it on the small speaker.

Headroom lives down here. Low frequencies carry most of the energy in a mix, so the bottom is usually what is driving a limiter hardest. If a master sounds squashed and nothing obvious is doing it, look at the bottom first. And keeping it close to mono below roughly 150 Hz matters for the same reason, which is a number you can actually read rather than a habit.

Should I tune my 808 to the key of the track?

If the key reading is solid, yes, a low part that is out of key reads as wrong rather than muddy. If the key reading hedges or names two candidates, do not take a note from it and tune by ear against the chords instead. A confident wrong tonic is worse here than no tonic, because the whole low end gets tuned to it. We wrote about which direction key detection tends to be wrong in.

Is the sub band the same as the 808?

No. The sub band is 20 to 60 Hz, a range. An 808 usually has energy there and also well above it. The band table describes ranges, not parts.

Does a loud sub band mean my low end is too loud?

It does not. The gate reads which band leads, not whether any band is excessive. Loudness is a separate reading.

Limits

The filename label is weak, as stated above, so the hit rates are comparative and not accuracy figures. What the gate does once it opens, and the rest of what gets measured while a set plays, is on the Ableton page. The corpus is one working sample library, which carries the taste of whoever built it, so the 2.4 percent base rate is a property of that library rather than of music. And everything after the gate opens is knowledge we act on, not a second measurement: we measured when to speak, not whether the advice worked.