LUFS Loudness Meter

Open an audio file in an editor and you see its peaks. The picture shows how close the signal comes to the ceiling of the format. It says very little about how loud the thing will sound.

A comparison published by Sound On Sound makes the gap concrete. One track, ripped from an early 1980s compact disc and wide in its dynamics, measured a true peak of -1.4 dBTP and an integrated loudness of -18.8 LUFS. Another, from a recent remastered collection, measured +0.5 dBTP and -7.8 LUFS. The two peaks are less than 2 dB apart. The two loudness figures are 11 LU apart, a disparity the author called painfully obvious on playing them back to back. A meter that watches peaks would call these recordings nearly equal. Anyone with ears would not.

Broadcasters lived with that mismatch for decades. European stations levelled programmes with the quasi-peak programme meter, whose integration time is 10 ms, against a permitted maximum that was usually -9 dBFS. The European Broadcasting Union’s production guidelines describe the result plainly: uniform peak levels of programmes, but widely varying loudness levels. The same guidelines call the jump in level at programme breaks a major source of irritation for audiences, and EBU R 128 names loudness inconsistency as the cause of the most viewer and listener complaints.

Music had its own version of the problem. Early compact disc mastering pushed signals to peak close to 0 dBFS, the reference level in effect became the clipping level, and producers learned to squeeze ever more of the signal up against that ceiling. The practice earned a name, the loudness war: a trend of increasing audio levels in recorded music that reduces fidelity. Lawmakers noticed the television half of it. In the United States the CALM Act became law on December 15, 2010, and told the Federal Communications Commission to make an industry practice on loudness, ATSC A/85, mandatory for commercials.

What the meter weighs

The measurement those rules rely on is Recommendation ITU-R BS.1770, first issued in 2006 and revised five times since. It has four stages: a frequency weighting, a mean square for each channel, a weighted sum across channels, and a gate. Nothing in it is exotic. That was the point.

The weighting is called K. Its first filter is a shelf that accounts for the acoustic effect of the head, modelled as a rigid sphere, and lifts everything above about 2 kHz by 4 dB. Its second is a high-pass filter, the revised low-frequency B-curve, which discounts deep bass. Then comes plain arithmetic. Each channel’s filtered signal is squared and averaged, the left and right surround channels are multiplied by 1.41, the low-frequency effects channel is left out, and the sum is turned into decibels. A constant of -0.691 is added, so that a full-scale 997 Hz sine on one front channel reads -3.01.

The unit has two names. The ITU writes LKFS, for loudness, K-weighted, relative to nominal full scale. The EBU writes LUFS and states that the two are equivalent. A loudness unit, LU, is the same size as a decibel: raise a signal by 1 dB and the reading rises by 1.

What differs from dBFS is the thing being counted. A sample-peak figure describes the largest sample. LUFS describes weighted average power. The EBU’s alignment check ties the two scales together at one point: a 1 kHz stereo sine wave peaking at -18 dBFS should read -18.0 LUFS. For a sine the numbers agree. For the two tracks above, peak and loudness differ by about 17 in one case and 8 in the other.

Was something this simple good enough? The ITU tested it against listeners who matched the loudness of 336 excerpts of broadcast and recorded material to a reference. The correlation between their judgments and the meter was 0.977. The recommendation notes that, for typical broadcast material, a simple energy-based measure proved about as robust as far more detailed perceptual models.

The gate decides what counts

An average over a whole programme has a flaw. Long quiet stretches pull the figure down, so a film with silences would measure lower than its dialogue sounds. The measurement is therefore gated. The audio is cut into blocks of 400 ms that overlap by 75 percent. Blocks quieter than -70 LKFS are discarded as silence. The loudness of what remains is computed, and every block more than 10 LU below that level is discarded as well. The final figure, the integrated loudness, comes only from the survivors, which is why the EBU describes it as a measure of foreground loudness.

You can watch the gate work. Press the Gating test button in the tool above. It builds one of the EBU’s compliance signals: a 1 kHz stereo tone for 10 seconds at -36 dBFS, 60 seconds at -23 dBFS, and 10 seconds at -36 dBFS again. Average the power over all 80 seconds and you get -24.2. The meter reads -23.0 LUFS, the value the EBU requires within 0.1 LU, because the quiet ends sit more than 10 LU under the ungated level and are dropped. In the 3D view the bars at both ends turn gray. The Reference tone button shows the plain case: 20 seconds at -23 dBFS reads -23.0 LUFS.

The gate is a convention, and conventions get revised. EBU R 128 first appeared in 2010 with a relative gate of -8 LU; in August 2011 it moved to -10 LU. Nor does everyone measure the same thing. R 128 asks for the whole signal, without emphasis on speech, music or effects. The American practice, A/85, is built around an anchor element, which is usually the dialogue.

Three more numbers

One figure cannot describe a programme, so meters report others. Momentary loudness uses a sliding window of 0.4 seconds and short-term loudness a window of 3 seconds, both without a gate. Their maxima show how hard the loudest moment hits. For adverts and promos the EBU caps short-term loudness at -18.0 LUFS, 5 LU above the target, to keep short items from being overly dynamic.

Loudness range, LRA, describes spread. It takes the short-term values, removes those below -70 LUFS and those more than 20 LU below the average of the rest, and reports the distance from the 10th to the 95th percentile. The percentiles are deliberate: a fade-out cannot stretch the low end, and a single gunshot cannot set the high end. The EBU’s own illustration is the film The Matrix, with a range of 25 LU. For anything shorter than a minute the figure is not recommended, because too few 3-second values exist.

True peak answers a different question: will this clip? Samples are snapshots, and the waveform between them can rise higher. BS.1770 describes a tone at a quarter of the sample rate, unluckily phased, reading 3 dB low on a sample-peak meter. Its remedy is to oversample by four, to 192 kHz for 48 kHz audio, and read the peak of the reconstructed signal in dBTP. The risk is practical. Lossy coding can create inter-sample peaks above the sample level, and Apple warns that compression may clip a signal that ignores its peak recommendation.

Hence the ceilings. R 128 allows -1 dBTP for linear audio in production. A/85 asks for peaks below -2 dB TP, and the EBU recommends -2 dBTP ahead of MPEG-1 Layer 2 and AC-3 coding.

What platforms do with the number

Streaming services normalize at playback. Spotify says it adjusts tracks to -14 dB LUFS according to the ITU standard, turning louder masters down and softer ones up. The upward move has a limit. Spotify leaves 1 dB of headroom, and its own example is a track at -20 dB LUFS with a true peak of -5 dB FS, which is lifted only to -16. Listeners can also choose a loud setting at -11 or a quiet one at -19. Apple Podcasts recommends an overall loudness around -16 dB LKFS, within 1 dB, with a true peak no higher than -1 dB FS.

Broadcast sits lower. R 128 sets -23.0 LUFS and A/85 sets -24 LKFS. The EBU’s streaming supplement accepts that online services typically use a higher target than broadcast, and gives -20 to -16 LUFS as an interim range for streams that broadcasters prepare themselves.

The consequence for anyone mastering is blunt. Where playback is normalized, pushing a master louder than the target gains nothing, because it is simply turned down; Sound On Sound called maximising loudness in that setting completely futile. The incentive has not vanished, though. Spotify notes that its web player and third-party devices such as TVs and speakers do not normalize, and one critic quoted by Wikipedia argues that the loudness war did not end but evolved.

What a loudness meter cannot do

A meter measures. It does not fix. The tool on this page reads a file and changes nothing in it; reaching a target still means applying gain, and often limiting, somewhere else. The number is also an estimate. BS.1770 says so itself: measured loudness involves uncertainty that depends on listeners, material and listening conditions, and the algorithm is not, in general, suitable for pure tones. It weighs the whole mix, so it cannot say whether the dialogue is audible under the music.

Mono needs care. A one-channel file measures about 3 LU lower than the same audio copied to two channels, and FFmpeg’s documentation notes that a mono file meant for playback on a stereo system gives a perceptually incorrect reading unless that is compensated. The tool offers the same correction as a switch.

True peak has two blind spots. A meter that oversamples by four can still under-read; the EBU puts the shortfall at about 0.5 dB and advises leaving a full decibel of headroom. And the peak that matters is the one after encoding, which no measurement of the master can show. Measure the encoded file as well.

None of this makes the meter less useful. It replaced a question nobody could answer by looking, how loud is this, with a number that tracks what people hear. Know that number before a platform applies its own.