The number every platform measures

The number every platform now measures

For twenty years, mastering meant making things as loud as possible, because on a playlist the louder track drew more attention. That competition (the loudness war) ended for practical purposes when streaming services began normalising playback. They now measure every track and turn it down to a common target.

That changes the arithmetic completely. If you squash a mix to be 6dB louder than the target, the platform turns it down 6dB, and all you have achieved is delivering a flatter, more fatiguing version at the same perceived volume as everybody else.

What LUFS measures

LUFS (Loudness Units relative to Full Scale) is a measurement designed to correlate with perceived loudness, standardised as ITU-R BS.1770 and adopted as EBU R128 in Europe and ATSC A/85 in the United States. LKFS is the same thing under a different name.

Two features make it different from a simple average:

K-weighting. The signal is filtered before measurement, with a high-shelf boost above about 2kHz and a high-pass below about 100Hz, approximating the ear’s greater sensitivity in the midrange. So a bass-heavy track measures quieter than its raw RMS suggests, which matches what you hear.

Gating. Passages more than 10LU below the running average are excluded from the integrated figure. Without this, a film with long quiet stretches would measure very low and be turned up until the loud scenes were painful. The gate makes the number describe the programme’s typical level rather than its average including silence.

Three readings are reported:

• Momentary: a 400ms window.
• Short-term: a 3-second window. The useful one while mixing.
• Integrated: the whole programme, gated. The one delivery specifications name.

Loudness range (LRA) describes the spread between quiet and loud passages. A heavily compressed pop master might read 3LU; a film mix might read 15 to 20LU.

The numbers, as they stand

These change, and any list is a snapshot. Check the platform’s own documentation before a delivery.

• Spotify, YouTube, Amazon Music: around −14 LUFS integrated.
• Apple Music: around −16 LUFS.
• Podcasts: −16 LUFS for stereo is the usual recommendation, with −19 sometimes specified for mono.
• European broadcast (EBU R128): −23 LUFS, with a tolerance of ±0.5 or ±1.
• US broadcast (ATSC A/85): −24 LKFS.

The broadcast figures are much lower because broadcast preserves dynamic range for programmes that contain both whispering and explosions, and because it is a legal requirement in several countries that advertisements do not jump in loudness.

True peak, and why −1 is not paranoia

Alongside the loudness target, specifications give a true peak ceiling, usually −1 dBTP and sometimes −2 for lossy delivery.

A sample-peak meter reports the highest stored sample. The analogue waveform reconstructed between samples can be higher than any individual sample (inter-sample peaks) and a lossy encoder’s output can overshoot further still, because encoding and decoding do not preserve the exact waveform. A file peaking at −0.1 dBFS can therefore clip in a listener’s decoder even though it never clipped in your session.

Leaving 1dB of true-peak headroom costs nothing audible and removes the problem.

Hitting the target

Mix to sound right first, measure second, and adjust with a gain change rather than by squashing.

If your mix measures −20 LUFS against a −16 target, raise it 4dB and check the true peak. If that would clip, you need a limiter, but only for the last couple of decibels. Reaching for 8dB of limiting to hit a number is where dynamics go.

# measure
ffmpeg -i mix.wav -af ebur128=peak=true -f null -

# two-pass normalisation to -16 LUFS with a -1.5 dBTP ceiling
ffmpeg -i mix.wav -af loudnorm=I=-16:TP=-1.5:LRA=11:print_format=json -f null -
ffmpeg -i mix.wav -af loudnorm=I=-16:TP=-1.5:LRA=11:measured_I=-20.3:measured_LRA=9.1:measured_TP=-3.2:measured_thresh=-30.9:linear=true out.wav

The two-pass form matters. In one pass, loudnorm works dynamically and can pump audibly. Measure first, feed the measurements back, and with linear=true it applies a single gain change, which is what you want.

Podcast and video specifics

For speech, the number to watch while mixing is short-term, not integrated. A podcast that averages −16 LUFS with one guest at −22 and another at −11 is a bad podcast with a correct number. Balance the speakers against each other first, then set the whole thing to the target.

For video, the dialogue is the anchor. Mix the dialogue to sit consistently, place music and effects relative to it, then measure the programme.

The free path

• ffmpeg’s ebur128 filter and loudnorm do the measurement and the correction, free, and are what many web tools wrap.
• Youlean Loudness Meter has a free version that is a full EBU R128 meter.
• Audacity includes a loudness normalisation effect with a LUFS target.
• Ardour has built-in loudness analysis and reports integrated LUFS and true peak per range.
• DaVinci Resolve’s free Fairlight page has a compliant loudness meter with presets for the broadcast standards.

The limitation

LUFS is a model of perception, calibrated on a particular set of programme material, and it does not always match what people hear. Two tracks at the same integrated LUFS can sound different in loudness, particularly when one is dense and the other sparse, or when their spectral balances differ substantially. K-weighting is a broad approximation of one equal-loudness contour, and the ear’s response changes with level.

And normalisation is not universal. It can be switched off by the listener on several services, applies differently to albums and to single tracks, and is applied inconsistently across the platforms that do it. So a delivery target is a sensible convention rather than a guarantee, and the real argument for respecting it is that over-compressing to beat it gains nothing and costs the mix.