Why your EP does not sit together

This is the failure that survives every check people actually run. Each song was mixed well, each was mastered to a sensible level, each measures fine. Then you listen to the record in order and track three is noticeably duller than track two, track five jumps out, and the whole thing feels assembled rather than made.

The reason nothing caught it is that every tool most people use measures one file at a time. The fault is a property of the relationship between files, and you cannot see a relationship by looking at one end of it.

Two kinds of drift

Loudness drift is the obvious one. Integrated loudness varies across the set — not by the amount the songs intend, but by an amount that comes from each track being finished on a different day in a different mood.

Some variation is correct and musical. A ballad between two loud songs should be quieter. What is not correct is the loud songs disagreeing with each other by two decibels for no reason anybody could articulate.

Tonal drift is the one people miss. Track one has a little more air, track four has a heavier low mid, track six is brighter because it was mastered after a long day when your ears were tired. Individually, all within normal range. In sequence, the listener experiences it as the record losing its footing — and they do not attribute it to tone, they attribute it to the songs being weaker.

Listeners do not hear a tonal inconsistency. They hear a record that stops being convincing halfway through.

Why normalisation does not save you

The common assumption is that per-track normalisation on streaming flattens everything out, so loudness consistency no longer matters. This is half true and the half that is false is the one that hurts.

On album playback, some platforms apply album normalisation — one gain offset for the whole record, preserving the relationships you built. Your intended dynamics between tracks survive, including the unintended ones.

On shuffle and playlist playback, per-track normalisation applies and the level differences do get flattened. But normalisation adjusts level, not tone. Tonal drift passes through it completely untouched, and so does any damage from over-limiting on the loudest tracks.

So the situation is: your level inconsistencies are hidden in some contexts and exposed in others, and your tonal inconsistencies are exposed in all of them. Neither is a reason to skip the check.

What to measure across the set

Four numbers per track, compared across the record rather than against a target.

• Integrated loudness. Look at the spread, not the values. A range of two to three LU across a set is normal and musical. Six is a record that was never sequenced.
• Loudness range. A track with a much lower LRA than its neighbours has been limited harder, and it will sound flat in context even when its level matches.
• True peak. Should be consistent across the set. One track at minus zero point three among files at minus one means one file went through a different chain.
• Spectral balance in broad bands. Low, low-mid, presence, air. You are looking for one track that sits away from the group, not for an ideal curve.

The spectral comparison is the one worth building a habit around, because it is the one your ears cannot do reliably. Comparing the tone of a track you heard four minutes ago against the one playing now is a memory task, and auditory memory is poor and gets worse as you get tired — which is exactly the state you are in when you sequence a record.

Fixing the outliers

The correct move is almost always to move the outlier toward the group, not to reprocess the group.

Find the two or three tracks that anchor the record — usually the ones you finished when you were sharpest, and usually the ones you like most. Treat those as the reference. Then bring the outliers to them with small broad moves: half a decibel of shelf, a decibel of level, nothing surgical.

Resist the urge to make everything identical. A record where every track has the same spectral balance is as wrong as one that wanders, just in a less obvious way. You are removing accidents, not variety.

Move the outliers to the record. Do not move the record to the outliers.

Listening to it properly

After the numbers, listen to the transitions and only the transitions. Last thirty seconds of one track into the first thirty of the next, in sequence, all the way through. This takes about six minutes for an EP and it is the single most useful listening pass in the whole process.

The transitions are where inconsistency becomes audible, because it is the only moment the listener has both tracks in immediate memory. If a record survives its transitions, it holds together.

And do it on one system, at one volume, in one sitting. The whole point is to remove the variation that came from working across many days. Introducing more of it at the checking stage defeats the exercise.