Your mix sounds muddy and the analyzer says it is fine
An analyzer reads level. Mud is a ratio. Here is the measurement that separates the two, the corpus its thresholds came from, and when it stays quiet.
Level and balance are two different measurements
Band level is the energy inside a frequency band, read in dBFS. Band balance is that band's share of the total energy, compared against the share a typical mix carries there. Most analyzers show you the first one. The complaint you are trying to solve is almost always the second one.
The gap is easiest to see by turning something down. Play a loop at -30 dBFS and every band on the meter drops with it, so an absolute threshold reads clean, in every band, every time. Nothing about the material changed. The proportions are identical, and proportions are what your ears are listening to.
Key takeaways
- Absolute band level moves with the fader. A band's share of the total does not, so the share is the number that survives a gain change.
- A share only becomes readable after two corrections: one for bandwidth, one for the fact that real music is not pink noise.
- Our upper thresholds are musical, not statistical. They sit where a clearly audible 6 dB EQ move would put the score.
- Our lower thresholds are statistical, because a thin band is a description, not a fault.
- When no single frequency stands out inside the band, we name none. The absence is the finding: there is no one culprit to cut.
Three layers turn a share into a number you can compare
Layer one, the raw share. The band's energy divided by the total energy between 20 Hz and 16 kHz. The ceiling is not arbitrary. Above 16 kHz, lossy coded material is dead, and putting a dead region in the denominator would move the air share according to where the file came from rather than what is in it.
Layer two, the pink difference. The mud band spans one octave, the air band about two thirds of one, so their raw shares differ by construction. The measured share is divided by the share an equal energy per octave spectrum (pink) would take in the same band, expressed in dB. Bandwidth drops out of the equation.
Layer three, the typical mix offset. Real music is not pink. Read raw, every track ever made would report low air. So a neutral offset is subtracted per band, and that offset is measured rather than chosen.
| band | window | neutral offset | high gate | low gate |
|---|---|---|---|---|
| mud | 200 to 400 Hz | -2.3 dB | +4.0 dB | -11.0 dB |
| harsh | 2 to 5 kHz | -2.1 dB | +3.5 dB | -8.0 dB |
| air | 10 to 16 kHz | -11.8 dB | +5.0 dB | -9.5 dB |
A score near 0 dB means the band sits where a typical mix has it. Positive is more than typical, negative is less.
The offsets came from a library, not from a preference
The corpus is a working sample library: 1008 wav files, of which 983 could be measured. The remaining 25 were too short or practically silent.
Raw percentiles over all of it do not give a usable gate, and the reason is worth stating out loud. A kick's harsh share is about -47 dB, and that is not a defect, that is a kick.
| subset | files | mud median | harsh median | air median |
|---|---|---|---|---|
| whole corpus | 983 | -4.7 | -4.8 | -17.8 |
| drum loops | 116 | -5.1 | -5.2 | -10.0 |
| vocals | 119 | +1.6 | -5.5 | -20.1 |
| hats and cymbals | 89 | -29.3 | +0.4 | +3.4 |
| kicks | 66 | -18.7 | -47.2 | -57.4 |
One sided tails like those pushed the low gate down toward -33 dB, which is a number that means nothing.
A mix balance judgement is only defined for broadband material, so the offsets were read from the broadband subset: files where none of the three bands is practically empty, each one sitting above 25 dB under the pink reference. That subset is 525 files, and the neutral offsets in the table above are its medians.
Two gates, two different logics, on purpose
The upper gate is musical. Taking the 90th percentile of the corpus was the obvious move and it would have been wrong: it defines harsh as harsher than the average of a sample library, and sample libraries are bright and hard by design. So the anchor is a question instead. How far does a known EQ move push the score? The same 120 broadband files were measured dry and processed.
| move | +3 dB EQ | +6 dB EQ |
|---|---|---|
| mud, 300 Hz, Q 1.0 | +1.97 dB | +3.77 dB |
| harsh, 3.2 kHz, Q 2.0 | +1.51 dB | +3.05 dB |
| air, 12 kHz, Q 0.7 | +2.66 dB | +5.28 dB |
The high gates are that last column, rounded: 4.0, 3.5, 5.0. Which gives the threshold a sentence a person can check: high means the band sits about one clearly audible 6 dB move above where a typical mix has it.
The lower gate is statistical. Muddy and harsh are complaint words, and the absence of a complaint is not a fault. A thin band is only worth mentioning when it is genuinely unusual, so the low gate is the 10th percentile of the broadband corpus: emptier than nine out of ten real files.
When "cut 200 to 350 Hz" is the wrong advice
Every cheat sheet names a frequency. Naming one is only honest when one is actually there, and the first version of our peak finder learned that the hard way.
It compared every twelfth of an octave against its own neighbourhood, one sixth of an octave either side. It failed its own test. A real harsh resonance sits around Q 2, roughly 0.7 octave wide, so the peak lifts its neighbours with it and no difference appears. An obvious bell at 3.2 kHz went straight past it.
Each region is now read against the band's own slope, a straight line in log frequency against dB, and the peak is the largest residual above that line. Two consequences, both of them wanted:
- A Q 2 bell leaves a bump above the line, so it gets caught.
- A whole band lifting evenly is absorbed into the line, so the band reads high and no frequency is named at all.
The second case is the one most advice gets wrong. If the entire low mid region is up, cutting 200 to 350 Hz does not remove mud, it removes body. Pointing at a culprit that is not there would be a lie, so the field is simply not written. Its absence is the information: there is no single offender, and the fix is an arrangement or a level decision rather than a notch.
The measurement had to prove itself before it was allowed to speak
Thirty one automated checks stand behind the three numbers, and they exist because a confident wrong reading is worse than no reading. The ones worth quoting:
- Pink noise reads within 1 dB of zero in all three bands, which is the ratio maths being right rather than approximately right.
- The same signal 24 dB quieter produces the same three scores and the same three levels.
- A 12 dB injection at 3.2 kHz reads harsh +7.4 and trips high. The same injection at 300 Hz reads mud +9.3.
- A clean reference reads mud +3.6, harsh +1.8, air -2.1, and stays inside the ok band in all three.
- A broad lift trips high and names no frequency.
What this measurement cannot tell you
The corpus is loops and stems, not finished mixes. On the broadband subset the gates fire on 25.5 percent of files for mud, 25.3 percent for harsh and 31.2 percent for air. That rate is honestly high, and it is left where it is for a reason: a solo synth stem legitimately carries a big harsh share, while a mixdown off your master bus sits much closer to the middle of the distribution.
The missing input is stated rather than buried. There is no finished mix corpus here yet, only eight songstarters, which is too few for statistics. If one arrives, the work is already defined: the neutral offsets move to its median, the high gates stay exactly where they are because they are musical rather than statistical, and the low gates settle on its tenth percentile.
Questions
Is this the same thing as a spectrum analyzer?
No. An analyzer draws level against frequency and leaves the interpretation to you. This is one number per band that already carries the comparison: how much of the total is here, versus how much is normally here. You can build the same judgement by eye from an analyzer, but you have to hold a reference in your head while you do it.
Does a high reading mean I have to cut something?
No. It means the band is carrying more of the total than a typical mix does. On a bright hat loop that is correct and expected. The reading is a description until you decide it is a problem, and the decision stays yours.
Does it work on a single sound rather than a mix?
The measurement works on anything with content in the band, but the judgement does not transfer. On a one shot, balance describes the sound rather than diagnosing it, which is why a kick reading very low harsh is not a finding.
Why 200 to 400 Hz for mud when other sources say 250 to 500?
Because this window is a diagnosis rather than a cut. A corrective EQ aims at where the problem is most audible, and a bit lower catches body buildup, which is what accumulates when many parts overlap. The window is a measurement boundary, not a suggested filter setting.
The short version
If a band reads high, something in the arrangement is putting more energy there than a typical mix carries. If a frequency is named, there is a resonance you can point at. If no frequency is named, do not go looking for one.
This is the measurement behind one of the things Miles says while you work, and the thresholds above are the actual numbers in the build, not an illustration. The rest of what it measures while a set plays, and the list of what it will never touch, is on the Ableton page.