Why this matters

If you run any video product — a streaming service, an OTT platform, a conferencing app, a telemedicine tool — the loudest complaint you will ever get is about loudness: the ad that blares, the next episode that whispers, the one speaker on the call nobody can hear. Fixing that starts with measuring it correctly, and most teams measure it wrong because they reach for a peak meter that has nothing to do with perceived loudness. This article is the foundation: it defines every term you need before you can read the platform-target tables, the normalization standards, and the true-peak limiting rules in the rest of this section. Read it once and the words "−14 LUFS, −1 dBTP" stop being jargon and become an instruction you can act on.

First, the decibel: it is a ratio, not a unit of loudness

Almost every confusion about loudness traces back to the decibel, so we start there. The decibel, written dB, is not a fixed amount of anything — it is a way of expressing how many times bigger one value is than another, on a compressed scale. It is a ratio, like "twice as much" or "a tenth as much", dressed up in logarithms so that very large and very small ratios fit on the same axis.

The reason audio uses a logarithm is that hearing is logarithmic. Going from a whisper to a normal voice and from a normal voice to a shout each feel like one "step up", even though the second step carries far more physical energy. A logarithm turns those equal-feeling steps into equal-sized numbers, which is exactly what we want from a loudness scale.

Because a decibel is a ratio, it always needs a reference — the "compared to what". When you see a suffix after dB, that suffix names the reference. dBFS means decibels relative to full scale, where 0 dBFS is the loudest sample a digital system can store and every real signal sits below it as a negative number. So −6 dBFS is half the maximum amplitude, −12 dBFS is a quarter, and so on. Here is the only piece of decibel arithmetic you need to keep:

a change of +6 dB  ≈ double the amplitude
a change of −6 dB  ≈ half the amplitude
a change of +20 dB = ten times the amplitude

Hold onto two ideas from this section. First, a decibel describes a ratio, never an absolute loudness. Second, the suffix (FS, TP, or the LUFS family below) tells you the reference the ratio is measured against. Every term in this article is a different answer to the question "decibels compared to what, and measured how".

Peak: the highest single sample, and why it lies about loudness

The simplest thing a meter can report is the peak: the value of the single loudest sample in a stretch of audio, expressed in dBFS. A peak meter answers one narrow question — how close did the signal come to the digital ceiling at 0 dBFS, beyond which samples "clip" and distort. That is a real and useful question. It is just not the same question as "how loud is this".

The reason peak fails as a loudness measure is that loudness is about sustained energy over time, while a peak is one instant. Consider two clips. The first is a steady, dense pop master that sits near the top of the scale for three full minutes. The second is a sparse acoustic recording that is quiet almost everywhere except for one sharp drum hit that briefly touches the same level. A peak meter reports the same number for both, because both reached the same single highest sample. Your ears report that the first clip is dramatically louder, because it delivers far more energy across the whole duration. Peak measured the spike; it never measured the sustained loudness that the ear actually integrates.

There is a subtler problem with peak that we flag now and cover in depth later. The "peak" a normal meter shows is the highest stored sample — but when a digital signal is reconstructed back into a continuous analog wave for your speakers, the smooth curve drawn between samples can rise higher than any individual sample. That hidden overshoot is the inter-sample peak, and the measurement that catches it is True Peak, in dBTP. We define it properly in the True Peak section below; for now, just remember that even the peak number you trust can under-read the real signal.

RMS: an average of energy, closer but still not loudness

If a single peak is too narrow, the natural fix is to average. RMS — short for root mean square — is a way of averaging a signal's energy over a window of time rather than reporting one instant. The name describes the recipe: square every sample (which makes negatives positive and weights big values heavily), take the mean of those squares over the window, then take the square root to get back to the original scale. The result is a number that tracks the sustained level of the audio, not its momentary spikes.

RMS is a real step toward loudness. Go back to the two clips above: RMS correctly reports the dense pop master as far higher than the sparse acoustic track, because it measures energy across the window instead of one sample. For decades RMS-style meters (VU meters and their descendants) were the best practical loudness indicator available, and they are still useful.

But RMS has one blind spot that the ear does not share: it treats every frequency as equally loud. Human hearing does not. We are most sensitive to the midrange — roughly the band where speech lives — and much less sensitive to very low bass and very high treble at the same physical energy. A track with heavy sub-bass can carry enormous RMS energy down low that you barely perceive as loudness, while a track with a strong vocal presence can feel loud at lower RMS. RMS measures the energy correctly; it just measures it as if your ear were flat, and your ear is not. Closing that gap is the whole job of the next measurement.

LUFS: loudness measured the way the ear hears it

This is the term the rest of the section is built on, so we define it carefully and in plain language first. LUFS stands for Loudness Units relative to Full Scale. It is a number, expressed on a decibel-style scale below a 0 ceiling (so real audio is negative, like −23 LUFS or −14 LUFS), that is engineered to match perceived loudness — how loud a human actually judges the audio to be — rather than its raw electrical energy.

LUFS is not a vendor's idea. It is defined by an international standard, ITU-R BS.1770, currently in revision 5 (published November 2023), titled Algorithms to measure audio programme loudness and true-peak audio level. When a streaming platform says "we normalize to −14 LUFS", it is pointing at a specific, reproducible algorithm from that document. The algorithm has three moving parts, and you can understand all three without any heavy math.

Part one: K-weighting — filter the audio like an ear

Before measuring energy, the algorithm passes the audio through a fixed two-stage filter called K-weighting, which reshapes the sound to approximate the ear's own frequency sensitivity. The filter does two simple things. It rolls off the deep bass with a high-pass filter (energy below about 80 Hz counts for less, because we perceive it as less loud), and it gives a gentle boost of roughly +4 dB to the high frequencies above about 2 kHz with a "high-shelf" filter, reflecting the slight emphasis the head and ear add up there. K-weighting is the step that fixes RMS's blind spot: after this filter, equal measured energy corresponds much more closely to equal perceived loudness.

Part two: mean energy — average across the program

After K-weighting, the algorithm measures the mean energy (a root-mean-square-style power average) of the filtered signal across the whole piece of audio. For multichannel audio it sums the channels, and here is the one number worth remembering: the surround channels are weighted slightly higher (a factor of about 1.41, roughly +1.5 dB each) because sound arriving from behind is perceived as a touch louder, while the low-frequency-effects (LFE) "point-one" channel is left out of the loudness sum entirely. The summed, weighted energy is converted to the LUFS scale with a fixed calibration offset built into the standard.

Part three: gating — ignore the silences

A long film is full of quiet passages and outright silence. If you averaged those into the loudness number, a dialogue-heavy drama with long pauses would measure artificially quiet, and you would over-boost it. So BS.1770 applies gating: it measures the audio in short overlapping blocks of 400 milliseconds and throws away the blocks that fall below two thresholds. First, an absolute gate discards anything below −70 LUFS — true silence and room tone. Then a relative gate discards anything more than 10 LU below the average of what remains — the merely quiet moments. The final integrated loudness is the average of only the blocks that survived gating: the parts a listener would actually call "the program".

The payoff of all three steps together: two clips with the identical peak and even similar RMS can land 8 LUFS apart, and the LUFS number will agree with which one your ears call louder. That is why the entire streaming and broadcast world standardized on it.

Three meters measuring the same waveform and disagreeing: a peak meter catches one tall transient, an RMS meter averages the sustained energy, and a LUFS meter adds K-weighting and gating to match perceived loudness Figure 1. The same audio, three measurements: peak reports the loudest instant, RMS reports average energy, and LUFS reports perceived loudness after filtering like an ear and ignoring the silences.

Left-to-right pipeline of the ITU-R BS.1770 loudness algorithm: input audio passes through K-weighting, then mean energy per channel, then gating, producing the integrated loudness in LUFS, with a plain-language caption under each stage Figure 2. The three real steps of the BS.1770 algorithm — filter like an ear, average the energy, drop the silent blocks — turn a raw waveform into one integrated LUFS number.

The vocabulary around LUFS: LU, LKFS, M, S, I

Four more words travel with LUFS, and they trip people up only because they are never defined. Here they are, each in one line.

LKFS (Loudness, K-weighted, relative to Full Scale) is the same number as LUFS — the same scale, the same algorithm. The two are interchangeable; North American broadcast documents (and the US CALM Act) write LKFS, while European documents write LUFS. If a spec says −24 LKFS and another says −24 LUFS, they mean the identical thing.

LU (Loudness Unit) is the relative cousin of LUFS. One LU equals one dB of loudness difference. You use LUFS for an absolute measurement ("this track is −16 LUFS") and LU for a difference or a tolerance ("turn it down 3 LU", "±1 LU allowed"). Same scale, but LUFS is anchored to full scale and LU is a gap between two loudnesses.

The last three are the time windows the standard meters report, and you read all three on a real loudness meter. Momentary (M) loudness is measured over the most recent 400 ms — the instantaneous feel. Short-term (S) loudness is measured over the most recent 3 seconds — useful for catching a section that drifts loud or quiet. Integrated (I) loudness is the single gated average over the entire program — this is the number you deliver against, the one a platform checks for compliance.

Term Full name What it measures When you use it
LUFS Loudness Units rel. to Full Scale Absolute perceived loudness Targets, delivery ("−14 LUFS")
LKFS Loudness, K-weighted, rel. to FS Identical to LUFS US / broadcast specs ("−24 LKFS")
LU Loudness Unit A loudness difference (1 LU = 1 dB) Tolerances, adjustments ("±1 LU")
Momentary (M) — Loudness over last 400 ms Live monitoring, instantaneous
Short-term (S) — Loudness over last 3 s Spotting section-to-section drift
Integrated (I) — Gated average over whole program Compliance, the delivered number

Loudness Range (LRA): how much the loudness moves

Integrated loudness gives you one number for a whole program, but it hides an important thing: how consistent that loudness is. A talk-radio show and a film with whispered dialogue and explosions can share the same integrated LUFS while feeling completely different to listen to. Loudness Range, abbreviated LRA and measured in LU, captures that spread — it describes how much the loudness varies across the program, from its quiet passages to its loud ones.

LRA is defined in EBU Tech 3342, a companion to the EBU R128 recommendation. The method measures short-term loudness across the whole program, builds a statistical distribution of those values, and reports the spread between roughly the 10th and 95th percentiles (after its own gating). A small LRA, perhaps 3–5 LU, means tightly controlled, consistent loudness — typical of a pop master or a podcast. A large LRA, 15 LU or more, means a wide dynamic journey — typical of a film score or a classical recording.

Why you care: LRA tells you whether a piece of audio will survive a loud, distracting environment. A film with a 20 LU range sounds magnificent in a quiet cinema and becomes unwatchable on a phone in a moving train, because the quiet dialogue disappears under the noise. Knowing the LRA before delivery tells you whether you need a separate, more compressed mix for mobile. It is a planning number, not a compliance number.

True Peak and dBTP: the overshoot a normal peak meter misses

We promised to return to this, and it matters enough to earn its own section. A standard peak meter reports the highest stored sample value. But audio is not played back as a staircase of discrete samples — a digital-to-analog converter reconstructs a smooth continuous wave that passes through those sample points. Between two samples, that reconstructed curve can rise higher than either sample on its own. That hidden overshoot is the inter-sample peak, and a sample-peak meter is blind to it.

True Peak, measured in dBTP (decibels True Peak), is the measurement that catches it. The method, also defined in ITU-R BS.1770, is to oversample the signal — typically by 4× — which means mathematically inserting interpolated points between the real samples to estimate the continuous wave, then reading the true maximum of that denser curve. The number it returns is often a few tenths of a dB higher than the sample peak, and occasionally more than a full dB higher.

This is not academic. A master that reads exactly 0.0 dBFS on a sample-peak meter can have true peaks at +0.7 dBTP or higher. When that file is decoded — especially after a lossy codec like AAC, which can add its own overshoot — those inter-sample peaks clip the analog stage of a phone or a pair of earbuds and you hear crackle. That is why streaming platforms set a True Peak ceiling below zero: the EBU R128 maximum permitted level is −1 dBTP, and most streaming services adopt −1 dBTP as well, leaving headroom for the overshoot the sample meter never showed. (The deep dive on inter-sample peaks and limiting lives in True peak, dBTP and the inter-sample peak problem.)

Waveform reconstruction showing sample dots that all sit below 0 dBFS while the smooth analog curve drawn through them rises above the ceiling, with the overshoot labelled as the inter-sample true peak Figure 3. Every stored sample sits under 0 dBFS, yet the reconstructed analog wave overshoots the ceiling between samples — the inter-sample peak that True Peak (dBTP) catches and sample peak misses.

Putting the numbers together: a worked reading

Let the four measurements speak together on one imaginary delivery. Suppose your loudness meter, run across a finished 30-second commercial, reports this:

Integrated loudness  = −18.0 LUFS
Loudness Range (LRA) =   4.0 LU
True Peak (max)      =  −0.4 dBTP
Sample peak (max)    =  −1.1 dBFS

Read it as a sentence. The commercial's overall perceived loudness is −18.0 LUFS, which is moderate — louder than the −23 LUFS broadcast norm, quieter than a −14 LUFS music-streaming target. Its loudness range of 4.0 LU is narrow, so it is consistently loud throughout, with no quiet passages that will vanish on a phone. But look at the peaks: the sample peak is a comfortable −1.1 dBFS, yet the True Peak is −0.4 dBTP — the inter-sample overshoot added 0.7 dB the sample meter never showed. If your delivery spec demands −1 dBTP, this file fails on true peak despite looking safe on the sample meter, and you would lower it or apply a true-peak limiter before delivery.

That single reading is the whole article in miniature: peak protects against clipping, True Peak protects against the hidden clipping, LUFS describes perceived loudness, and LRA describes how much that loudness moves.

A common mistake: "I normalized to 0 dB, so it's as loud as possible"

The single most frequent loudness error we see is treating peak normalization as loudness normalization. Peak normalization raises a file until its loudest sample just touches 0 dBFS (or some ceiling). It guarantees you used the full digital range — and it tells you almost nothing about how loud the result feels. Two files both peak-normalized to 0 dBFS can be 10 LUFS apart in perceived loudness, because one is dense and sustained and the other is sparse with one tall transient.

The consequence shows up the moment your audio meets a modern platform. Spotify, YouTube, Apple Music, and broadcast chains do not care about your peak — they measure your integrated LUFS and turn the whole track up or down to hit their target. A track you "maximized" to 0 dBFS but that sits at −9 LUFS will simply be turned down by the platform to match a −14 LUFS target, and all the clipping you risked to get loud buys you nothing. The fix is to mix and master to a LUFS target and a dBTP ceiling, not to a peak. Loudness is the goal; peak is only a safety rail.

Where Fora Soft fits in

In the products we have built since 2005 — video streaming and OTT platforms, e-learning, telemedicine, and live conferencing — loudness is rarely a "nice to have"; it is where users notice the seams. On streaming and OTT work we measure every asset's integrated LUFS and True Peak before it enters the library, so an episode, an ad, and a trailer do not jump in volume when they play back to back. In real-time conferencing and telemedicine, the same vocabulary drives the automatic-gain and leveling stages that keep a soft-spoken patient and a loud clinician at a comfortable, even level. The recurring lesson across all of them is the one this article teaches: measure perceived loudness with LUFS and guard the ceiling with dBTP, and the "why is this so loud / so quiet" tickets mostly disappear.

What to read next

Call to action

References

  1. ITU-R BS.1770-5, Algorithms to measure audio programme loudness and true-peak audio level (November 2023). The controlling specification for LUFS/LKFS: the two-stage K-weighting filter, the multichannel energy sum with surround weighting (~+1.5 dB) and LFE excluded, the 400 ms gating blocks with 75% overlap, the −70 LKFS absolute gate and −10 LU relative gate, and the 4× oversampled True Peak (dBTP) measurement. https://www.itu.int/rec/R-REC-BS.1770/
  2. EBU R128 v4.0, Loudness normalisation and permitted maximum level of audio signals (June 2020). Defines the −23 LUFS programme target, the ±0.5 LU tolerance (±1 LU for live), and the −1 dBTP maximum permitted true-peak level that this article cites for the streaming ceiling. https://tech.ebu.ch/docs/r/r128.pdf
  3. EBU Tech 3341, Loudness Metering: 'EBU Mode' metering to supplement EBU R128 loudness normalisation. Source for the Momentary (400 ms), Short-term (3 s), and Integrated meter definitions used in the vocabulary table. https://tech.ebu.ch/docs/tech/tech3341.pdf
  4. EBU Tech 3342, Loudness Range: A measure to supplement EBU R128 loudness normalisation. Defines Loudness Range (LRA) in LU as the statistical spread of short-term loudness across a programme. https://tech.ebu.ch/docs/tech/tech3342.pdf
  5. ITU-R BS.1771-1, Requirements for loudness and true-peak indicating meters. Companion to BS.1770 specifying how compliant loudness and true-peak meters must behave. https://www.itu.int/rec/R-REC-BS.1771/
  6. ATSC A/85:2013, Techniques for Establishing and Maintaining Audio Loudness for Digital Television (the "CALM Act" recommended practice). Source for the North American −24 LKFS practice and the LKFS terminology, distinguished here from the EBU's −23 LUFS. https://www.atsc.org/atsc-documents/a85-techniques-for-establishing-and-maintaining-audio-loudness-for-digital-television/
  7. AES, Loudness Project — Learn More: Peak Metering and Loudness. First-party engineering material from the Audio Engineering Society on why sample-peak metering under-reads inter-sample peaks and why true-peak metering is required. https://aes2.org/resources/audio-topics/loudness-project/learn-more/
  8. Lund, Thomas, ITU-R BS.1770 Revisited (TC Electronic). Engineering paper on the rationale and behaviour of the BS.1770 gating and K-weighting design, used here for the plain-language explanation of why gating exists. http://magnetic.beep.pl/Loudness/Lund_BS1770_revisited.pdf

Note on source hierarchy (per editorial policy): every numeric claim about the LUFS algorithm (K-weighting shape, gating thresholds, block length, surround weighting, 4× true-peak oversampling) is taken from the controlling standard ITU-R BS.1770-5 (2023), not from secondary summaries. Where popular articles conflate the −23 LUFS EBU broadcast target with a universal "standard", this article follows the specs and distinguishes EBU R128's −23 LUFS from ATSC A/85's −24 LKFS. Platform-specific targets (−14 LUFS etc.) are named only as orientation here and are covered with primary citations in article 3.4.