Loudness Normalisation: A Technical Reference for Musicians

What loudness should my master be?

For music released to streaming services, set a true-peak ceiling of −1 dBTP and let integrated loudness land where the music puts it rather than forcing it to a number. If you want a published figure to aim at, AES TD1008 recommends −16 LUFS for track-normalised music; Spotify normalises playback to −14 LUFS. Most contemporary commercial masters measure somewhere between −14 and −9 LUFS, but that is an observation about current practice, not a specification, and no service asks you to land in it. Spotify asks for true peak below −1 dBTP, or below −2 dBTP if your master is louder than −14 LUFS. Every major service turns loud masters down on playback, so a master at −6 LUFS is not played louder than one at −14 LUFS — it is played at the same level with less dynamic range surviving. The only thing that changes with excessive limiting is what you lose. Choose a ceiling of −1 dBTP, measure integrated loudness on the finished master, and stop pushing when transients start to flatten. There is no single correct number, but there is a correct method.

Why loudness normalisation makes over-limiting self-defeating

The argument is arithmetic, not aesthetics.

A streaming service that normalises playback measures your track's integrated loudness, compares it to a target, and applies a fixed gain to the whole file at playback time. If Spotify's target is −14 LUFS and your master measures −14 LUFS, the applied gain is 0 dB. If your master measures −8 LUFS, the applied gain is −6 dB. If it measures −11 LUFS, the applied gain is −3 dB.

Now compare two masters of the same mix. Master A is finished at −14 LUFS integrated with peaks at −1 dBTP, giving a peak-to-loudness ratio of 13 LU. Master B is pushed to −8 LUFS integrated with the same −1 dBTP ceiling, a PLR of 7 LU. On playback at a −14 LUFS target, Master B is attenuated by 6 dB. Its peaks now sit at −7 dBTP. Its loudness is identical to Master A's. The 6 dB of crest factor that limiting removed is simply gone, and nothing in the playback chain gives it back.

Loudness normalisation converts a loudness war into a dynamic range war that the loudest master automatically loses. That is the whole case, and it does not depend on taste.

The corollary is more useful than the warning. Because the service sets the playback level, the level of your master is no longer a competitive variable. What remains competitive is everything level was previously spent on: transient definition, the impact of a snare against a sustained pad, the difference between a verse and a chorus, the sense that a vocal is in front of the band rather than pressed into it. Under normalisation, your master's peak-to-loudness ratio is the only part of your loudness decision the listener can actually hear.

Two qualifications. Normalisation is defeatable — listeners can switch it off, and some contexts (DJ software, a sync placement, a downloaded file played from a car aux input) never apply it. And upward normalisation is not unconditional. Spotify states it applies positive gain to softer masters but "consider[s] the headroom of the track, and leave[s] 1 dB headroom for lossy encodings." A very quiet master with high peaks may therefore not be raised all the way to the target. This is the real reason not to master at −24 LUFS "because it will be turned up": it may not be.

What LUFS actually measures

LUFS means Loudness Units relative to Full Scale. LKFS (Loudness, K-weighted, relative to Full Scale) is the same unit under a different name; ITU documents use LKFS, EBU documents use LUFS, and they are numerically identical. LU is the same size as a dB — one Loudness Unit equals one decibel — but LUFS is an absolute scale and LU is a relative one. A change of "+3 LU" and "+3 dB" mean the same magnitude; "−14 LUFS" and "−14 dB" do not mean the same thing at all.

The measurement defined in ITU-R BS.1770 is not RMS and not peak. It is a mean square of a filtered signal, summed across channels with per-channel weighting, then gated in time. Three things distinguish it from a simple RMS meter: the K-weighting filter, the channel summation, and the gating. The filter and the gating are where almost all misunderstanding lives.

K-weighting: two filter stages

K-weighting is a cascade of exactly two second-order sections applied to every channel before the mean square is taken.

Stage 1 — the "head" filter. A high-shelf that lifts the top of the spectrum (published 48 kHz coefficients: b0 = 1.53512485958697, b1 = −2.69169618940638, b2 = 1.19839281085285, a1 = −1.69065929318241, a2 = 0.73248077421585). It models the acoustic effect of a head in a sound field — the shadowing and diffraction that make a human listener more sensitive to high frequencies arriving from the front than a bare microphone would be.

Stage 2 — the RLB high-pass. Revised Low-frequency B-curve: a high-pass that rolls off the bottom, reflecting that low frequencies contribute less to perceived loudness than their energy suggests. This is why a track with enormous sub-bass content does not measure as loud as it feels in a club and does measure quieter than an engineer expects.

The gain of the K-weighting curve at 1 kHz is +0.698 dB, a linear factor of 1.0836. This is not an arbitrary constant; it falls out of the shelf and high-pass responses at that frequency. The practical consequence of the whole chain is that a full-scale 1 kHz sine applied to a single channel of a stereo pair reads −3.01 LKFS, exactly as BS.1770-5 states — the −3.01 being the one-of-two-channels sum, not the filter.

Why the curve exists at all

An unweighted RMS meter says a 40 Hz sine and a 3 kHz sine at the same amplitude are equally loud. No listener agrees. K-weighting is a deliberately crude approximation of equal-loudness behaviour — far simpler than the Fletcher-Munson curves, and chosen to be simple so that independent implementations agree to a fraction of a LU. K-weighting is not an attempt to model hearing accurately; it is an attempt to model it identically across every meter in the world, and that is why it works. Two conformant meters measuring the same file agree closely enough to make cross-service comparison meaningful.

Which edition applies

BS.1770 has six editions (2006, 2007, 2011, 2012, 2015, 2023). BS.1770-4 (October 2015) is the edition most deployed meters cite; BS.1770-5 (November 2023) is the edition currently in force. Every mechanic described in this document — the two filter stages, the 400 ms blocks, both gates, the 4× oversampling floor — is verified against BS.1770-5 as published.

The gating mechanism, in full

Gating is the part of the specification most often described incorrectly, including in otherwise careful sources. Integrated loudness is a doubly gated measurement.

Step 1: block the signal

The signal is divided into 400 ms gating blocks with 75% overlap. A 75% overlap on a 400 ms block means a new block starts every 100 ms. Each block gets its own K-weighted mean-square loudness value. The overlap is not decorative: it prevents a loud event that straddles a block boundary from being split and under-counted.

Step 2: the absolute gate

Any block measuring below −70 LKFS is discarded and takes no further part in the measurement. This removes digital silence, fade tails, room tone and the gaps between songs, so that a track with a long silent outro does not measure quieter than the same track trimmed.

Step 3: the relative gate — the part that is usually stated wrongly

Compute the mean loudness of the blocks that survived the absolute gate. Subtract 10. That result is the relative threshold. Discard every block below it. The integrated loudness is the mean of the blocks that survive both gates.

The distinction matters and is frequently lost: the relative gate is set 10 LU below the absolute-gated mean, not 10 LU below the ungated mean of the whole file. BS.1770-5 is explicit that the relative threshold is derived by "subtracting 10 from the result" of the absolute-gating measurement. If you compute the mean over all blocks including the sub-−70 LKFS ones, you get a lower mean, a lower threshold, more quiet blocks admitted, and an integrated value that reads too low — and the error grows with the amount of silence in the file. This is the single most common reason two meters disagree on the same track.

The audible consequence of the −10 LU relative gate is worth internalising: the integrated loudness of your track is effectively the average loudness of its loud parts, not of the track as a whole. A song with a whispered 40-second intro and a wall-of-sound chorus is measured almost entirely on its choruses. This is why adding a quiet intro to a finished master barely changes its integrated reading, and why engineers who expect it to are surprised.

Momentary, short-term, integrated: which meter, when

Three time windows are standardised, and they answer three different questions. EBU Tech 3341 defines the meter behaviours.

Momentary (M) — 400 ms, ungated. The same window as a gating block. Momentary loudness tracks individual events: a snare, a vocal consonant, a downbeat. Use it for catching things — a kick that spikes 6 LU above everything around it, a section that jumps out. Do not use it to make level decisions; it is far too twitchy, and chasing a momentary meter produces exactly the over-compressed result normalisation punishes.

Short-term (S) — 3 s, ungated. Three seconds is roughly a musical phrase. Short-term loudness is the right meter for balance decisions inside a track, because three seconds is the timescale on which listeners actually perceive a section as loud or quiet. Use short-term to compare a verse against a chorus, to check that a bridge does not collapse, and to see the shape of your arrangement as a number.

Integrated (I) — whole programme, doubly gated. This is the number services normalise against and the only one that belongs in a delivery spec. Measure it over the entire track from first sample to last, not over a selected section. Integrated loudness is a delivery specification, not a mixing tool; if you are watching it while you work, you are watching the wrong meter.

A practical rule: mix and master by short-term, verify by integrated, investigate anomalies with momentary.

Loudness Range (LRA) and the −20 LU gate

Loudness Range, specified in EBU Tech 3342, is a single number describing how much a track's loudness varies over time. It is computed from a distribution of short-term (3 s) loudness values: LRA is the difference between the estimates of the 10th and 95th percentiles of that distribution, which is why brief extremes at either end do not dominate the result.

LRA has its own gating, and it is not the same as the integrated gate. Tech 3342 sets the LRA relative threshold at −20 LU below the absolute-gated loudness level, not −10 LU. The wider gate is deliberate: LRA is trying to characterise variation, so it must admit quieter passages that the integrated measurement is designed to exclude.

This is a live implementation bug, not a theoretical one. Meters that reuse the integrated −10 LU relative gate for LRA report values that are systematically too small, because they have thrown away exactly the quiet material that constitutes the range. If your meter reports an LRA noticeably lower than another meter on the same file, the first thing to suspect is a −10 LU gate where −20 LU belongs. You can test this yourself: append 20 seconds of material 15 LU below the body of the track. A correct implementation's LRA rises substantially; a broken one barely moves.

As orientation rather than targets: heavily limited electronic and pop masters commonly land around 3–5 LU, a well-controlled full-band mix around 6–9 LU, and orchestral and acoustic recordings frequently above 12 LU. LRA is descriptive; pursuing a particular LRA number is as misguided as pursuing a particular LUFS number.

True peak versus sample peak

The difference

A sample peak meter reports the largest absolute sample value in the file. That is not the largest value the signal reaches: samples are points on a continuous waveform, and the waveform between two samples can rise above both. The reconstructed analogue waveform can exceed the highest sample in the file, and a sample-peak meter is structurally incapable of seeing it. These excursions are inter-sample peaks; the level of the reconstructed waveform is true peak, in dBTP.

A digitally "safe" file whose samples top out at −0.1 dBFS can have true peaks well above 0 dBTP. Nothing is clipped in the file. Everything downstream that reconstructs the waveform — a DAC, a sample-rate converter, a lossy decoder — may clip it.

Why oversampling is required

To see a true peak you must reconstruct the waveform between the samples. BS.1770 specifies doing this by oversampling: a minimum of 4× oversampling at 48 kHz, with the recommendation noting that "Higher sampling rates and over-sampling ratios are preferred." 4× is the conformance floor, not the accurate answer; at 4× the measurement can still under-read genuine peaks by a few tenths of a dB. Use 8× or 16× oversampling if your meter offers it; the additional CPU cost is trivial and the additional accuracy is not. A meter without an explicit true-peak or dBTP mode is a sample-peak meter, whatever its manual implies.

Why lossy encoding raises peaks

A lossy codec — AAC, Ogg Vorbis, MP3 — does not reproduce the waveform sample-for-sample. It reproduces a perceptually similar one, and the decoder's output routinely overshoots the encoder's input. The overshoot is not a defect but an unavoidable consequence of quantising and discarding spectral detail, and it scales with how hard the source was limited. A master at exactly 0.0 dBTP decodes with peaks above 0 dBFS, and the decoder clips them.

Every published spec that addresses this says the same thing. AES TD1008: "For all content, it is recommended that the Maximum True Peak level not exceed −1 dBTP at the codec input of lossy-encoded streams." Spotify: keep true peak "below −1dB TP (True Peak) max", and for masters louder than −14 LUFS, "keep True Peak below −2dB to avoid extra distortion."

What ceiling to actually use

Set your limiter's true-peak ceiling to −1.0 dBTP for normal masters, and −2.0 dBTP if your master is louder than −14 LUFS integrated. That is not a margin of superstition; both figures are published requirements from the sources above, and the second exists precisely because harder-limited material overshoots more in the codec.

Two footnotes. A ceiling of −0.1 dBTP is a delivery error for streaming, not a stylistic choice. And a true-peak limiter set to −1.0 dBTP still needs to be a true-peak limiter: many limiters' ceiling controls are sample-peak, and setting one to −1.0 gives you inter-sample peaks somewhere above −0.5 dBTP.

If you want to check a finished file, mazufa.com hosts a free browser-based loudness and true-peak checker that runs entirely on your own device and uploads nothing.

What each service actually publishes

The distinction between published and widely reported is the most important thing in this section, and almost no article on this topic maintains it.

Published specifications

Spotify. Target: −14 LUFS integrated. Spotify states it adjusts tracks "to −14 dB LUFS" and measures "according to the ITU 1770 standard." True-peak ceiling: below −1 dBTP; below −2 dBTP for masters louder than −14 LUFS. Spotify also states that "positive gain is applied to softer masters so the loudness level is −14 dB LUFS," qualified by headroom: "We consider the headroom of the track, and leave 1 dB headroom for lossy encodings to preserve audio quality." On album normalisation, Spotify notes the album is normalised as a unit so that "the softer tracks are as soft as you intend them to be" — the relative levels of tracks within an album are preserved.

EBU R 128. Target: −23 LUFS. This is a broadcast recommendation, not a streaming-music target, and is included because it is routinely misapplied: a music master at −23 LUFS is not more compliant with anything, and — because upward normalisation is headroom-limited — may play back quieter than intended.

AES TD1008 (Recommendations for Loudness of Internet Audio Streaming and On-Demand Distribution). This is the most useful public document for musicians, and the most misquoted.

• Track-normalised music: −16 LUFS, tolerance +0.2 LU.
• Album normalisation, loudest track: −14 LUFS, tolerance +0.2 LU.
• Speech-led content: −18 LUFS, tolerance +1 LU, measured as dialogue integrated loudness.
• Maximum true peak: −1 dBTP at the codec input for lossy-encoded streams.

The −18 LUFS figure in TD1008 is the target for speech-led content — news, talk, drama — and quoting it as the music target is a common and serious error. The music figure is −16 LUFS. If you see −18 LUFS presented as an AES music recommendation, the source has misread the document.

TD1008 also makes an explicit dynamics argument worth quoting: "A recording with high peak to loudness ratio (PLR) is often perceived as clearer and less fatiguing than one that has been excessively peak-limited."

Services that publish no normalisation target

Apple Music, YouTube Music, Amazon Music, TIDAL and Deezer do not publish a normalisation target. They normalise — that is observable — but the specific figure is not a published specification, and any article that presents one as such is presenting a measurement or an inference as a document.

The figures commonly circulated are: Apple ≈ −16 LUFS, YouTube Music ≈ −14 LUFS, Amazon Music ≈ −14 LUFS, TIDAL ≈ −14 LUFS, Deezer ≈ −15 LUFS. These are widely reported but not published by the services; treat them as third-party observations of a moving target, not as specifications. They come from independent measurement, they have changed over time, and no service has committed to them. Do not compute an expected playback gain from them, and do not master to one.

The practical consequence is smaller than it looks. Every reported figure sits between −16 and −14 LUFS, which is a 2 LU spread. No master should be re-cut over 2 LU.

Practical guidance

How to decide a target

Work backwards from the music, not forwards from a number.

1. Fix the ceiling first. −1 dBTP, true-peak, oversampled. This is non-negotiable and is the only genuinely hard constraint in the list.
2. Master for the music. Get the balance, tone and dynamics right with the limiter doing as little as possible. Do not look at the integrated meter during this stage.
3. Measure the result. Whatever integrated loudness you have arrived at, that is your candidate.
4. Sanity-check the range. If your integrated value is between roughly −14 and −9 LUFS, you are inside the range where every published target and every reported figure lives within a few LU, and no service will do anything dramatic to your track. If you are louder than −9 LUFS, ask what the limiting bought you, because playback normalisation will take the level back.
5. Only then consider adjusting, and adjust by changing the mastering, not by adding limiting.

Targeting a LUFS number by adding limiting until the meter reads correctly is the exact behaviour normalisation was designed to make pointless. If you need the number to move, move it with gain staging, with mix decisions, or with less compression — not with more.

Genre matters in a way a single number cannot capture. Dense electronic, metal and modern pop masters commonly land between −10 and −8 LUFS because the material is inherently sustained; jazz, folk and orchestral masters land between −18 and −13 LUFS because it is not. Both are correct. The mistake is a folk record limited to −8 LUFS, not a folk record measuring −16 LUFS.

If your track is already limited

You have a finished, heavily limited master and you have just read the above. Three honest options, in order of preference:

Remaster from the mix. The only real fix. Peak limiting is not reversible; the transients that were removed are not in the file. If you have the mix, back the limiter off and re-render.

Lower the ceiling and leave the rest. If you only have the master and its true peak is above −1 dBTP, apply true-peak limiting or clip-safe gain reduction to bring it to −1 dBTP and stop there. This fixes the delivery error without pretending to fix the dynamics.

Do nothing. A master at −8 LUFS with a −1 dBTP ceiling is not broken. It plays back at the same level as everything else, slightly flatter than it could be — a real cost, but smaller than that of a rushed remaster. Over-limiting is a missed opportunity, not a defect — the file will play, it will simply not play as well as it could have.

Dynamics versus level

Keep these separate in your head, because the conflation is the source of most bad decisions.

Level is where the whole track sits. It is a single number, it is set by a fader, it is free, and under normalisation it is chosen by the service, not by you.

Dynamics is the relationship between the loud and quiet parts, both moment to moment (crest factor, PLR) and section to section (LRA). It costs something to create, it is destroyed by limiting, and it cannot be restored downstream.

You are not competing on level any more — the service sets that — so every decibel you spend on level is spent out of a budget the listener never sees. Loudness normalisation did not remove your ability to make a loud-sounding record. It removed the ability to make one by being loud. Perceived intensity now comes from arrangement density, transient contrast, low-mid weight, and the size of the jump between sections — the things that survive a −6 dB playback gain unchanged.

What genuinely does not matter

• Hitting a LUFS target exactly. Nothing rounds, nothing fails, and the difference between −14.0 and −13.4 LUFS is inaudible. A tolerance of ±1 LU is generous and no one will notice it.
• Different masters for different services. The published and reported targets span roughly 2 LU. One master at −1 dBTP with sensible dynamics is correct everywhere. Service-specific masters are a solution to a problem that normalisation already solved.
• Sample rate and dither on a 24-bit master. Deliver at the mix's native rate; upsampling adds nothing, and 24-bit dither sits far below anything that survives distribution.
• The −23 LUFS broadcast figure. Unless you are delivering to broadcast, R 128 is not your specification.
• Metering plugin brand. Any BS.1770-conformant meter with true-peak mode and adequate oversampling agrees with any other to within a fraction of a LU. Disagreements trace to gating or oversampling implementation, not to quality.

Summary

Measure integrated loudness on the finished master with a BS.1770-conformant meter. Set a true-peak ceiling of −1 dBTP, or −2 dBTP if you are louder than −14 LUFS. Use short-term loudness for balance decisions and integrated only for verification. Remember that the integrated relative gate is −10 LU below the absolute-gated mean, and that LRA's is −20 LU. The service, not you, sets the playback level — spend the effort you would have spent on level on what the listener can still hear.

Mazufa distributes music with no upload fee, no subscription and no per-release charge; the only deduction is 5% of royalties received.

Sources

ITU-R BS.1770 — measurement algorithm, K-weighting, gating, true peak

• Recommendation ITU-R BS.1770-5 (11/2023), Algorithms to measure audio programme loudness and true-peak audio level (full text PDF): https://www.itu.int/dms_pubrec/itu-r/rec/bs/R-REC-BS.1770-5-202311-I!!PDF-E.pdf
• BS.1770 recommendation page and edition history (BS.1770-0 through BS.1770-5): https://www.itu.int/rec/R-REC-BS.1770/en

EBU — broadcast target, meter windows, Loudness Range

• EBU R 128, Loudness normalisation and permitted maximum level of audio signals (−23 LUFS): https://tech.ebu.ch/publications/r128
• EBU Tech 3341, Loudness Metering: EBU Mode metering to supplement EBU R 128 (momentary 400 ms, short-term 3 s): https://tech.ebu.ch/publications/tech3341
• EBU Tech 3342, Loudness Range: A measure to supplement EBU R 128 loudness normalisation (−20 LU gate; 10th/95th percentiles), PDF: https://tech.ebu.ch/docs/tech/tech3342.pdf

AES — streaming recommendations

• AES TD1008.1.21-9, Recommendations for Loudness of Internet Audio Streaming and On-Demand Distribution (−16 LUFS music, −14 LUFS album, −18 LUFS speech, −1 dBTP), PDF: https://aes2.org/wp-content/uploads/2024/01/20210924_TD1008_v3.13.pdf
• AES TD1008 document page: https://aes.org/technical-council/technical-document-aestd1008/

Spotify — published normalisation specification

• Spotify for Artists, Loudness normalization (−14 LUFS; −1 dBTP / −2 dBTP; positive gain and headroom; album normalisation): https://support.spotify.com/us/artists/article/loudness-normalization/

Not published by any service. The figures given for Apple Music, YouTube Music, Amazon Music, TIDAL and Deezer are widely reported third-party measurements. No primary source is cited for them because none exists; these services publish no normalisation specification.