The loudness war

Why did every record label mastering louder leave listeners with music nobody enjoyed more?

Between the late 1980s and the late 2000s, commercial records got steadily louder. Not louder as in played at a higher volume — louder as in mastered so that the average level sat closer and closer to the format's ceiling, with the peaks squashed down to make room. By the mid-2000s some releases had so little variation between their quiet and loud moments that the waveform was a solid block.

Nobody set out to do this. No listener asked for it. Many of the engineers doing it disliked it. And at the end of it, records were no louder in the only sense that matters, because you can turn a knob. So we have an outcome that almost every participant regretted, produced by people acting sensibly. That combination is the signature of a particular kind of problem, and it is worth identifying precisely which one.

Start with the fact the whole thing rests on. Louder sounds better. Not metaphorically — in controlled comparison, a version of a recording that is even half a decibel louder is preferred over the same recording quieter, by trained and untrained listeners alike, and they do not report hearing a difference in level. They report hearing more presence, more detail, more impact. This is why serious listening tests level-match their material; without matching, the test measures gain rather than quality.

Now ask what follows if you are a mastering engineer in 1998 and you know this. Your record is going to be heard next to other records — on the radio, in a shop, on a compilation, in a listening room where an A and R executive flips between candidates. In every one of those settings the comparison is immediate and the level is not matched. What is your best move?

Louder. Obviously louder. And notice this is not cynicism or laziness; it is a correct response to the environment. A quieter master genuinely loses the comparison it is going to be put in.

But now run the same reasoning for every other engineer, because they all know the same thing. Everyone raises their level. And here the structure gives itself away: loudness in this game is purely relative. The advantage exists only in the gap between your record and the next one. When everybody closes the gap, the advantage is exactly zero — and it is zero at a new, permanently higher average level.

So what was actually spent to buy nothing? Headroom. To raise average level without exceeding the ceiling you must reduce the distance between the loudest moments and the rest — heavy compression and limiting, and in the worst cases outright clipping of the peaks. That distance is dynamic range, and it is what carries the punch of a snare, the swell of a chorus after a quiet verse, the sense that the music is doing something. Push far enough and you get audible distortion; push moderately and you get a record that is tiring after twenty minutes for reasons a listener cannot name.

The canonical illustration arrived in 2008 with Metallica's Death Magnetic, whose mastering drew enough complaint to become a public argument — helped by the fact that the audio prepared for the Guitar Hero game had noticeably more dynamic range, so listeners could hear the same performances both ways and say which they preferred. They preferred the quieter one.

Which tells us the payoffs precisely. Each participant gains from being louder than the field. The field collectively loses by all being louder. No individual can improve matters by unilaterally stopping — a lone holdout just sounds weak next to everyone else. That is the exact shape of a collective-action problem: rational individual choices aggregating into an outcome every individual would trade away.

And that diagnosis predicts the cure. If the payoff comes from an unmatched comparison, remove the unmatched comparison. Broadcasters did this first, adopting loudness measurement standardised as ITU-R BS.1770 and the European practice EBU R 128, which normalises programmes to a target level. Streaming services then did the same for music: playback is normalised toward a target loudness — Spotify's is around minus fourteen LUFS — so a heavily squashed master is simply turned down to sit alongside everything else, and arrives with its dynamics spent and no advantage bought. The incentive did not have to be argued out of existence. It had to be metered away.

Three refinements.

First, compression is not a villain. Deliberate limiting is a legitimate and often essential creative tool, and plenty of loud records are excellent. The problem was never the technique; it was the technique being applied to win a comparison rather than to serve the music.

Second, the war is not simply "over". Normalisation removes the reward for a hyper-loud master, and average released dynamic range has recovered somewhat since the peak, but engineers still deliver loud masters out of habit, client demand, or concern about how a track sounds where normalisation is switched off. Incentives shift practice slowly.

Third, and most usefully: notice that the fix did not require anyone to become more virtuous or better informed. Everybody already knew. What changed was the comparison environment. When a bad outcome comes from a payoff structure, appeals to conscience tend to fail and referees tend to work — which is the transferable lesson here, and the reason this story keeps getting told outside audio engineering.

• How LUFS loudness measurement actually models the ear, and why peak level was the wrong meter.
• Whether normalisation has changed how records are arranged, not just how they are mastered.
• Other places where a relative advantage is bought with an irreversible cost.

Every term the collection defines is gathered in the glossary.