LUFS and true peak: what the meter is actually measuring

A peak meter tells you how close a recording came to clipping. It says nothing about how loud it sounds, which is why two masters that both peak at exactly −1 dBFS can differ by 8 LU in perceived level. LUFS exists to close that gap, and every streaming platform and broadcaster now normalises against it.

The measurement is completely specified — ITU-R BS.1770 for the loudness, EBU Tech 3342 for the range, and Annex 2 of the same recommendation for true peak — so it can be implemented rather than estimated. The analyser below does that, and reports the intermediate quantities that explain the number.

The chain, in order

1. K-weighting. Every channel is filtered by a high shelf followed by a high-pass. The result approximates the equal-loudness curve: content below 100 Hz and above 2 kHz counts for less than the midrange, so a bass-heavy master does not score as loudly as its level suggests.
2. Mean square in 400 ms blocks, with a 100 ms hop, summed across channels with the channel weights.
3. Gating. Blocks below −70 LUFS are discarded, then blocks more than 10 LU below the average of what remains are discarded too.
4. Average and convert, with the −0.691 dB offset.

The K-weighting response is worth seeing, because it explains most disagreements between a LUFS reading and an RMS reading:

Frequency	K-weighting gain	What it means
20 Hz	−13.28 dB	Subsonic content barely counts
38.1 Hz	−5.97 dB	The high-pass corner
100 Hz	−1.13 dB	Bass counts a little less than the mids
1 kHz	+0.70 dB	The reference frequency
2 kHz	+3.07 dB	Where the ear is most sensitive
10 kHz	+4.04 dB	The shelf has reached its plateau

Those numbers are measured from the standard's own coefficients, and they reveal the two stages exactly: stage 1 is a high shelf at 1500 Hz, +4 dB, Q 0.6, and stage 2 is a high-pass at 38.14 Hz with Q 0.5003. The high-pass parameters fall straight out of the published denominator; the shelf parameters were found by fitting, and reproduce the published response to 0.0023 dB across 20 Hz to 20 kHz.

Why the offset is −0.691

An arbitrary-looking constant usually cancels something, and this one cancels the K-weighting gain at 1 kHz, which is +0.698 dB. The two almost exactly cancel, and the consequence is the anchor that makes the whole scale meaningful:

A 1 kHz sine at −20 dBFS peak, in both channels of a stereo signal, measures exactly −20 LUFS — −19.99 LUFS with the coefficients above, the 0.01 dB being rounding. In mono the same tone measures −23.00 LUFS, which is within 0.01 dB of its own RMS level of −23.01 dBFS. So for a tone at the reference frequency, LUFS reads the level, and everything else is measured against that.

That also gives the fastest sanity check on any loudness implementation: if a −20 dBFS 1 kHz stereo pair does not read −20.0 LUFS, something in the weighting or the channel summation is wrong.

The gate, which is the part people skip

Without gating, a track with a quiet intro measures several LU lower than the same track without it, and the number depends on how much silence the editor left at the end. The gate removes that dependence in two stages. Load the two-level programme in the tool — 70% at −20 dBFS, 30% at −40 dBFS — and the measurement reports:

• Integrated loudness −23.05 LUFS, against −23.00 LUFS for the loud section alone. The quiet part has been removed entirely, not merely down-weighted.
• The relative gate lands at −34.5 LUFS, which is 10 LU below the ungated average. The quiet blocks at −43 LUFS are below it and never counted.
• 140 of 197 blocks survive. The other 57 are the quiet section and the transitions.

The two gates have different jobs. The absolute gate at −70 LUFS discards digital silence and dither; without it, a long fade-out would drag the average down and the relative gate would then be computed from a meaningless number. The relative gate removes material that is quiet relative to the rest of the programme, which is what makes the measurement independent of arrangement — a quiet verse and a loud chorus give the same integrated loudness whether or not the verse is there at all.

The exception is a programme shorter than 400 ms, and a file with nothing above −70 LUFS: both measure as −∞, which is the honest answer rather than a number that looks plausible.

Loudness range uses a different gate

LRA, defined in EBU Tech 3342, is the 10th to 95th percentile of the 3-second short-term loudness — but gated at 20 LU below the average rather than 10. That single difference has a visible consequence in the same two-level programme: the integrated measurement throws the quiet section away, while the range keeps it, because it sits within the wider gate. The tool reports LRA 20.000 LU, the exact size of the level step, from −43.00 to −23.00 LUFS.

So LRA and integrated loudness answer different questions. Integrated loudness is "how loud is this programme"; LRA is "how much does the level move". A film soundtrack might measure −27 LUFS with an LRA of 18 LU; a heavily limited pop master might measure −9 LUFS with an LRA of 3. The first is not quieter in any meaningful sense — the dialogue is normalised against the same scale — it simply reserves its level for the moments that need it.

True peak: why every sample can be below full scale

Digital samples are points on a continuous waveform, and the waveform between the points can go higher than any of them. A sine at a quarter of the sample rate, sampled 45 degrees off its crest, has every sample at −3.01 dBFS while the waveform reaches full scale — a 3 dB error that no sample-peak meter can see. Scale that signal so the samples sit exactly at full scale and the waveform reaches +3.11 dBTP, 13% above the highest representable sample.

That is what true peak measures, and it is why broadcast specifications are written in dBTP rather than dBFS: a converter fed a signal above full scale clips it, and the clipping happens in a device downstream of your meter. BS.1770 Annex 2 specifies at least 4x oversampling for the measurement, because at 4x the worst-case error for real programme material is under 0.1 dB, and the standard's own tolerance for these meters is ±0.5 dB.

The practical consequence for normalisation is worth stating explicitly. If content measuring −9 LUFS is turned down by 5 dB to reach a −14 LUFS target, its true peak falls by 5 dB too — and material that was already at 0 dBTP is then at −5 dBTP, which is fine. The failure happens the other way round: normalising a quiet, high-LRA programme up to a target and pushing its true peak past full scale. Check the peak after the gain, not before.

What LUFS does not measure

• Quality. It is a level measurement. A well-mixed track and a badly mixed one at the same LUFS score the same, which is the point.
• Peaks. LUFS integrates over 400 ms or more; a single sample clipping at +6 dBTP barely moves it. Loudness and peak are separate specifications for that reason.
• Mono compatibility, or the mono fold-down level. A mono file played through both speakers is 3 LU louder than the same file played through one, and both numbers are correct — they are different signals.
• Perceived quality of the weighting. K-weighting is a two-stage approximation to the equal-loudness contours, and it does not adapt to level the way the ear does. It is a standard, not a model of hearing.
• Anything about frequency balance. Two masters 3 dB apart in high-frequency tilt with the same LUFS will still sound different.

Five mistakes worth avoiding

1. Treating LUFS as a peak or RMS meter. It is a gated, weighted, block-based average, and none of those three steps is optional.
2. Comparing a mono measurement with a stereo one. The difference is exactly 3.01 LU, and it is the most common source of "my meter disagrees with yours".
3. Skipping the gate. An implementation without gating will disagree with every commercial meter on any programme with quiet passages, and the size of the disagreement depends on the arrangement.
4. Assume LUFS and LKFS are different units. They are identical; LKFS is the ITU's name and LUFS is the EBU's, and R128 uses them interchangeably.
5. Normalising to a target without checking true peak. Turning a quiet programme up is the case that clips.

Summary

Filter with K-weighting, average the power in 400 ms blocks, discard everything below −70 LUFS and everything 10 LU below the rest, and add −0.691 dB. Check the result against the anchor — a −20 dBFS 1 kHz stereo sine must read −20.0 LUFS — and measure true peak separately, because an "at −1 dBFS" master can be 3 dB above full scale between its samples.

Values shown are engineering aids rather than measurements; see the disclaimer and the tool index.