What -14 LUFS actually is
LUFS is a loudness unit from the EBU R128 recommendation and its ITU BS.1770 measurement standard. It is integrated across the whole track with a frequency weighting that approximates the ear, and it is designed so that two different masters of the same song measure the same number even if they sound nothing alike. That property is exactly why a streaming service can turn one down and another up and be confident about it.
So -14 LUFS is not a target for your mastering chain. It is a behaviour of somebody else's player. If your track arrives at -9 LUFS, a normalising service applies about 5 dB of gain reduction and plays it quieter relative to its neighbours, with no change to the audio beyond that single constant multiplication. Nothing about the quality of your master survives or dies at the number, which was chosen to make playback levels consistent across a catalogue rather than because a track at -14 sounds better than the same track at -11.
| Measurement | Typical value | What it tells you |
|---|---|---|
| Integrated loudness | -14 LUFS | Common normalisation target |
| Integrated loudness | -11 LUFS | No attenuation on most services |
| Loudness range | 2-5 LU | Spread between quiet and loud parts |
| True peak | -1 dBTP | Codec-safe ceiling |
| Sample peak | -0.1 dBFS | Not sufficient before lossy delivery |
Crest factor and why dynamics read as quality
Crest factor is the ratio of peak level to average level, and it is the most useful number nobody measures in a loudness war argument. A classical recording might run 18 dB. A modern loud master of the same music might run 9 dB. Both measure the same integrated loudness if the second is simply turned up, and the difference between them is the entire range that has been removed.
Listeners do not read crest factor, but they react to what it produces. Dynamics give a passage somewhere to arrive from. Flatten a mix and the loudness stays the same while the track becomes harder to follow, because there is no longer any relationship between quiet sections and loud ones.
- Measure loudness range rather than just integrated loudness; a 1 LU reading usually means the limiter has removed the arc of the track.
- Watch crest factor on a few genres as a reference and stop compressing once it drops below what that genre normally runs at.
- Compare at matched levels, because a louder file will always seem better in an A/B test and that result is an artefact of the test.
The trade-off is explicit: dynamic range costs average level, and average level is the closest thing to a quality signal a listener has. Giving up some of it buys protection at encoding time and headroom in the mix, which is a good bargain up to a point and a bad one past it.
Why codecs punish heavy limiting
Every lossy codec used for streaming works roughly the same way. It looks at a window of audio, predicts part of it from what came before, and then spends its limited bit budget on the error between the prediction and the truth. The cleverer the prediction, the more of the bit budget can be spent on detail rather than on repair. Now consider what a limiter does. It takes a waveform with a wide spread between peaks and average level, and pushes everything up against a ceiling. What comes out is dense and transient-heavy, with no repeating structure for a predictor to latch onto. The prediction fails exactly where the music is most detailed, which is the worst possible outcome for a fixed bit budget.
This is the technical reason the loudness war does not survive contact with a streaming catalogue. The extra loudness was bought with dynamics, and dynamics were partly what the encoder needed to do its job well. You are not only losing range, you are degrading the signal at the point where it is easiest to degrade.
The mitigation is boring and effective. Master with true peak under minus 1 dBTP rather than sample peak, so the reconstruction filter a decoder applies cannot push the signal past zero. Avoid brickwall limiting, and leave room in the low end, since heavily limited bass is the first thing to distort into mush at a low bitrate.
A chain that survives the round trip
A practical streaming master is shorter than most people's signal chains suggest: corrective EQ on anything genuinely wrong, a gentle multiband for balance, a transparent compressor if the mix needs glue, and a limiter used as a safety rather than as a level tool. The order matters more than the choice of unit.
Set the limiter's ceiling to minus 1 dBTP and enable true-peak detection so it responds to inter-sample overshoot rather than to individual samples. Aim the loudness reduction at the loudest passages rather than the integrated figure, which keeps quiet sections where the artist put them. Then stop, export at 32-bit float, and let the service handle the rest.
- Integrated loudness under minus 14 LUFS is a safe place to sit if you want zero attenuation on most platforms.
- Loudness range above about 2 LU means the track still has an arc.
- True peak at or below minus 1 dBTP means the reconstruction step has somewhere to go.
- Encode a short preview yourself at 96 kbps to hear what the codec will do to the busiest passage. It is the fastest honest check available and costs nothing.
The bottom line
Treat -14 LUFS as a specification written by somebody else's software, and master to the music instead. A track delivered at -14 LUFS with real dynamics survives lossy encoding better than the same track pushed to -6, and listeners have never reliably preferred the louder one. Keep true peaks under -1 dBTP, watch the crest factor, and stop when the song stops improving.
Check what mastering actually changes
432Hz MASTER moves the entire spectrum by one ratio. Match a reference level first, and every difference you hear after that is dynamics, not tuning.
Frequently asked questions
Should I master to exactly -14 LUFS?
Only if you have no better plan. It is a common normalisation target, so a louder master simply gets turned down on playback, and a quieter one gets turned up. Aim for a level that suits the material and let the platform handle the rest.
What is the difference between peak and true peak?
Sample peak reads the stored samples. True peak reconstructs the waveform between them and catches inter-sample overshoots that exceed 0 dBFS without any single sample doing so. Codecs reconstruct the signal the same way, which is why true peak is the number that matters.
Does louder always sound better?
Up to a point on the same system at moderate levels, then no. Listener preference plateaus, and past that the usual result is a worse master rather than a better-liked one, because the extra level has to come out of somewhere.
Why do codecs struggle with heavily limited material?
A codec predicts part of the waveform and codes the error, spending its bits on what it can predict. A flattened, dense waveform is mostly transients with no repeating structure, so the prediction is poor and the bit budget has to cover everything, which is where the artefact becomes audible.