Loudness, LUFS and streaming audio normalization

Why peak says nothing about how loud something sounds

dBFS (decibels relative to full scale) is an amplitude measure: 0 dBFS is the loudest sample the format can represent, and everything else is negative. It is a logarithmic scale, like almost any decibel relation, the same ratio-between-quantities idea covered in the unit conversion guide. The trouble is that peak only looks at one instant: the single loudest sample in the file. It has no idea whether that peak lasted a click or the whole record, nor whether the energy is concentrated in a high whistle or spread across the entire spectrum. The ear cares about exactly those two things.

A worked example lays the difference bare. A 997 Hz sine tone hitting 0 dBFS, fed to a single channel (left, centre or right), measures exactly −3.01 LKFS, that is how BS.1770 calibrates itself. Put the same tone in both channels of a stereo pair and the weighted sum doubles: the meter now reads 0.0 LUFS, with the peak untouched. Now take a dense pop master that also peaks at 0 dBFS: it integrates far louder, because it fills the whole spectrum and sustains that energy for minutes. Same peak, completely different loudness. Normalizing by the loudest sample equalizes the number nobody hears and leaves unequal the thing everybody hears.

What LUFS actually measures

The recipe lives in ITU-R BS.1770, the recommendation that defines how to measure loudness and true peak, now at revision BS.1770-5, November 2023, but stable in essence since the first version in 2006. The first step is K-weighting, a two-stage filter that approximates how we hear: a shelving pre-filter that lifts the highs by about +4 dB, modelling the effect of the head, followed by an RLB high-pass filter that cuts the very low frequencies. It is not the perfect audibility curve, but it is simple, reproducible and good enough to govern all professional audio. If you want to feel how differently bass and treble weigh on perception, play with the bass and treble equalizer.

After the filter, the algorithm squares the signal and averages it (the mean square) over 400 ms windows with 75% overlap. Each channel is measured separately and summed with a weight: this per-channel weighted sum is the heart of the measure.

The −0.691 constant is not magic: it cancels the K-filter gain at 997 Hz, anchoring the scale. The larger weights for the surround channels (1.41, about +1.5 dB) reflect that sound arriving from behind is perceived as slightly louder. Before going on, it helps to pin down the vocabulary, below, each acronym with its exact meaning.

dBFS

Decibels relative to digital full scale. 0 dBFS is the loudest representable sample; everything below is negative. Measures sample amplitude, not loudness.

dBTP

Decibels true peak: the peak of the reconstructed signal between samples, estimated by oversampling. It can exceed 0 dBFS even with no sample at 0.

LUFS

Loudness Units relative to Full Scale: the absolute BS.1770 loudness scale (K-weighting + gating).

LKFS

Loudness, K-weighted, relative to Full Scale: the name ATSC and ITU give to the SAME scale as LUFS. 1 LKFS = 1 LUFS.

LU

Loudness Unit: the same magnitude as 1 dB, but used relatively, the distance between two loudness values or to a target.

LRA

Loudness Range: the loudness variation of the programme (95th percentile minus 10th of the short-term blocks), in LU. High = dynamic; low = compressed.

gating

Discarding silent stretches before computing integrated loudness: an absolute gate at −70 LUFS and a relative gate 10 LU below the already-gated level.

Momentary / Short-term / Integrated

The three EBU Mode windows: momentary (400 ms), short-term (3 s) and integrated (the whole programme, gated).

The three windows, gating and LRA

A serious loudness meter shows three numbers, defined by EBU Tech 3341 as "EBU Mode". Each answers a different question about the same track.

• Momentary (M): a 400 ms window. Catches instantaneous loudness spikes, useful for spotting a one-off overshoot.
• Short-term (S): a 3 s window. Follows the mix as you work, tracking how the music evolves.
• Integrated (I): the whole programme, gated. This is the number you compare to a target like −14 or −23 LUFS.

Gating exists because a silence or a long pause would drag the average down and make a track look quieter than it sounds. BS.1770 applies two gates before computing integrated loudness: an absolute gate at −70 LUFS, which throws away what is essentially silence, and a relative gate placed 10 LU below the level already measured with the absolute gate. Only the blocks above that second gate enter the final average. Without gating, a podcast with long stretches of quiet speech and a shouted commercial could register the same integrated value, which would defeat the whole measure.

LRA (Loudness Range), defined in EBU Tech 3342, sums up how dynamic the track is: it takes the short-term blocks and measures the distance between the 10th and 95th percentiles, ignoring the extremes. It is a distribution measure, not an average, the same statistical logic of looking at the spread of the data that appears in the correlation and regression guide. A high LRA means strong contrasts between quiet and intense passages; a low LRA betrays heavy compression, where everything was flattened to the same volume.

True peak: the peak the sample meter never sees

Sample peak is the loudest sample in the file. True peak (dBTP) is the highest value the signal reaches once reconstructed as a continuous waveform, and the two almost never coincide. Between two samples that sit just under 0 dBFS, the curve the converter draws can rise higher than any existing sample. That is the inter-sample peak, and it is treacherous precisely because it is invisible to an ordinary peak meter. Seeing the waveform helps you understand where it approaches the ceiling: you can generate and inspect that in the waveform generator.

The broadcast standards: one ruler, two rules

BS.1770 is only the ruler, how to measure. The rules of how much are built on top of it. In Europe, EBU R 128 fixed the target at −23 integrated LUFS for programmes, with a ±0.5 LU tolerance (±1 LU for less predictable live material) and a maximum true peak of −1 dBTP. In the United States, ATSC A/85, which the CALM Act made mandatory for commercials, uses −24 LKFS, with a ±2 LU margin. Worth burning in: LKFS and LUFS are the same scale under different names; the 1 LU gap between the two targets is real, but small.

1. 2006ITU-R BS.1770

   The ITU publishes the algorithm defining K-weighting and true peak, the common ruler for the whole industry.

2. 2010EBU R 128

   The European Broadcasting Union sets the −23 LUFS target and the EBU Mode (Momentary, Short-term, Integrated, LRA).

3. Dec. 2010CALM Act signed

   In the US, the law (Public Law 111-311) directs the FCC to require loudness control of commercials.

4. Dec. 2012FCC rules in force

   The CALM Act rules take effect, adopting ATSC A/85 (−24 LKFS) as the mandatory reference.

5. 2010sStreaming adopts normalization

   One by one, music and video platforms begin normalizing playback by loudness, each with its own target.

6. 2023BS.1770-5

   The latest revision of the algorithm keeps the core intact and refines measurement details.

Streaming and the end of the loudness war

Streaming borrowed the idea from broadcast, but with a higher target. Spotify normalizes playback to −14 LUFS in Normal mode, with −11 LUFS (Loud) and −19 LUFS (Quiet) options for subscribers. Important: it does not alter your file, it applies a gain only at play time, down on loud tracks and up on quiet ones. And it respects true peak: per the official documentation, a track at −20 LUFS whose true peak sits at −5 dBFS is only raised to −16 LUFS, not −14, because +6 dB would push the peak to +1 dBFS and clip. The platform leaves 1 dB of headroom and stops short.

Here is how the loudness war ends. For decades, masters were compressed and limited ever louder to "win" on radio and CD: whoever played louder grabbed more attention. Loudness normalization retired that logic. Suppose you master at −6 LUFS, brickwalled; when it plays on Spotify in Normal mode, the track takes −8 dB of gain to drop to the very same −14 LUFS as a dynamic master sitting at −14. The two arrive equally loud, but yours lost its transients and its breathing room on the way. Squeezing no longer buys loudness; it only costs quality.

When every track is leveled to the same LUFS, squeezing the dynamics buys no loudness at all, it only costs the transients.

Communities like the Dynamic Range Database and campaigns like "Turn Me Up!" had been documenting this loss of dynamics long before the platforms acted. Today the economic incentive has flipped: a master that keeps its dynamics sounds equally loud as the rest and better in quality. The practice comes down to mixing with your ears, checking the integrated value and true peak at the end, and not mastering to a number louder than the destination target. Close the production loop with other audio tools, like the BPM detector.

• Measure the track’s integrated loudness (LUFS), not just the peak, use the volume normalizer or a meter with EBU Mode.
• Cap the true peak at −1 dBTP (−2 dBTP if the master is loud), with a true-peak limiter, not just a sample-peak one.
• Do not master above the destination target hoping for "loudness": platforms turn the track back down to their target.
• Preserve dynamics (watch the LRA): heavy compression gives no advantage after normalization.
• Check the whole album together, many platforms normalize per album to keep the relationships between tracks.
• Keep a more dynamic master as the source; it is easier to turn down later than to recover lost transients.