What normalisation actually does
A streaming service measures your track’s integrated loudness, compares it to its own target, and applies a gain change on playback. Nothing is re-encoded and nothing is re-compressed — it is a volume adjustment, applied to the whole track, decided once at ingest and applied every time the track plays.
The consequence is the part people find counter-intuitive: mastering louder than the target does not make you louder than anyone else. It makes the platform turn you down by the difference, and you arrive at the same playback level as everybody else, having spent dynamic range to get there.
The measurement is the integrated LUFS reading defined in ITU-R BS.1770 — the whole track, gated so that silence and quiet passages do not drag the figure down. If that sentence needs unpacking, what LUFS actually measures is the place to start; this article assumes it and gets on with what the platforms do.
The published targets
These are the figures each platform publishes, with the document each comes from. They move occasionally, so it is worth checking the source rather than trusting a number you learned once.
- Spotify — -14 LUFS integrated, ceiling -1 dBTP. Spotify Loudness Normalization documentation.
- Apple Music — -16 LUFS integrated, ceiling -1 dBTP. Apple Digital Masters technical specification.
- YouTube — -14 LUFS integrated, ceiling -1 dBTP. YouTube loudness normalisation, as measured and widely reported.
- Broadcast (EBU R128) — -23 LUFS integrated, ceiling -1 dBTP. EBU R128.
Two services can target different numbers and still be doing the same thing. Apple Music at −16 is not “quieter” than Spotify at −14 in any way you can act on — each one sets its own playback level, and your master is adjusted to each independently. The same file lands on target on both. Each of the two has a page of its own for the specifics: LUFS for Spotify and LUFS for Apple Music.
How loud should my master be?
Not −14 LUFS because Spotify says −14 — that is the most common misreading of the table above. The target describes where the platform will put your track, not where it wants you to deliver it. A master delivered at −14 and a master delivered at −9 both play at −14 on a normalised stream; the only difference the listener hears is the dynamics the louder one gave up.
So the material decides. A sparse, dynamic record is fine sitting under every target, because the platform raises it and the peaks are far from the ceiling. A dense record can be mastered well above the targets without losing anything it had, and is simply turned down. The mistake in either direction is chasing a number: compressing a folk record to reach −14, or squashing an electronic one past the point where the transients stopped being transients, for a loudness advantage that is removed on playback.
The practical version: master the record so that it sounds finished and holds up next to releases you respect in the same genre; then measure it, and act on the true peak. The integrated figure is information. The ceiling is a rule.
What if you are under?
It depends on the platform, and this is the one place the services genuinely differ. YouTube only turns tracks down: a master under its target plays at the level it was delivered. Apple’s Sound Check is documented the same way — it attenuates what is over −16 LUFS, and nothing published describes it raising what is under. Spotify does raise quiet tracks, and how it does so is where the ceiling matters.
Raising a track by 4 dB raises its peaks by 4 dB too. Spotify’s own documentation says it leaves 1 dB of headroom for the lossy encode when it applies positive gain, so a track at −20 LUFS whose true peak is already at −5 dBTP is lifted only to −16, not to −14 — and only the “loud” listening setting adds a limiter to push further. A quiet master with peaks near full scale is therefore heard quieter than everything around it, on the default setting, with nothing to rescue it. The same document asks that masters louder than −14 LUFS keep their true peak under −2 dBTP, because they are about to be turned down and re-encoded.
The conclusion is the same on every service: peaks well below the ceiling cost nothing, and they are the one thing that gives a normaliser room to do the simple thing.
Normalisation on or off — and album mode
Normalisation is a listener setting on most services — Spotify offers quiet, normal and loud levels, and Apple calls it Sound Check, which is the more consequential of the two because its default state varies rather than being on. Some listeners switch it off entirely. A track played back raw is heard at whatever loudness it was mastered to, sitting between other tracks at whatever loudness they were mastered to. This is the one scenario where a louder master is genuinely louder — and it is also the scenario where a master that leaned on the normaliser to sound right does not.
Album mode is the other setting worth understanding. When a listener plays an album through, the services that offer it normalise the album as a whole rather than each track, so that a deliberately quiet interlude stays quiet relative to the track after it. In shuffle or in a playlist, tracks are normalised individually. The consequence for a mastering engineer is that the level relationships across a record survive in album playback and are flattened in playlists, which is why a track that must work on its own is mastered to work on its own.
The one number worth being strict about
Inter-sample peak. Every one of these platforms delivers lossy audio, and a lossy encoder reconstructs a waveform that passes through your samples rather than stopping at them — so a master peaking at −0.1 dBFS on a sample-peak meter can exceed full scale in the encode and distort.
That is why the published ceilings are around −1 dBTP rather than 0. It is the one place where the guidance is a genuine warning rather than a preference, and it is measured differently from the peak meter in most DAWs: true peak oversamples the signal to find the curve between the samples. A limiter with a true-peak ceiling set a decibel under full scale is the whole fix, and it costs nothing audible. Distortion already baked into the file is a separate diagnosis — the clipping detector is what answers that one.
Podcasts and spoken word
Speech is normalised too, and usually to a different figure: Apple’s podcast delivery specification asks for −16 LUFS with a −1 dBTP ceiling, which is where most spoken-word platforms land. Speech also tends to be far more dynamic than music — one speaker sitting closer to the microphone than another can be several LU apart within one episode — so the loudness figure for a podcast is only half the job. Evening out the speakers is the other half, and no normaliser does it for you. What we do for spoken word covers that path, which is mostly a restoration and levelling problem rather than a mastering one.
Checking a master before upload
Measure the finished file — after the limiter, after dither, after any sample-rate conversion — rather than the session, and measure the whole track rather than the loudest section. Any BS.1770 meter gives the same integrated figure to within a tenth of a LU; the loudness checker on this site runs the measurement in your browser without uploading the file and sets the reading against the targets above.
Then read the result the right way round. “3 LU over Spotify” means the track will be turned down by 3 dB, which is not a fault. “0.4 dB over the ceiling” is the line that needs acting on.
What to actually do
Master for the song, check the numbers afterwards, and only act on the true peak. If your track lands at −8 LUFS because that is what it needs to be, the platforms will turn it down and it will sound like itself. If it lands at −8 LUFS because you were chasing loudness, you have given away dynamics for a volume nobody will hear.
If you would rather have the target set from the material than from a table, Euphona’s AI mastering chooses a loudness and true-peak target for the track it is given — a sparse record and a dense one get different answers — and reports every move it made, so the reasoning can be read back rather than taken on trust. The guide to mastering a song covers the same decisions by hand.
Keep reading