Sound in the Streaming Era — What Loudness Normalisation Does

Part of the loudness difference you hear is a platform rule, not a musician's choice

Related

Dynamic MarkingsWhat Is a Musical Sound

"This album is noisier than that one" — on streaming services, part of that statement is not the musician's decision but the platform's.

Where the problem comes from

Before digital, level differences between recordings of different eras were enormous:

• early twentieth-century discs were made very loud so they would cut through noisy playback equipment;

• 1980s mastering (the "loudness war") made things louder and louder;

• while classical recordings kept wide dynamics, because live performance is like that.

Streaming broke on this: played as-is, a 1970s recording is nearly inaudible while a 1990s master bulldozes everything else.

What platforms do

The platform's answer is automatic normalisation: analyse the whole file's loudness before playback, then bring it to a target level.

Two consequence chains follow:

Consequence one: old recordings get louder

The oratorio in the second example is exactly this: on a phone in 2026 it is far louder than it ever was on a hi-fi. Dynamics get squashed, because it must be pulled to the same level as everything else.

Consequence two: new recordings get quieter

In reverse, a track mastered to the modern spec is turned down by normalisation, because it already sat above the target.

Here is the paradox: the more carefully a version is mastered, the quieter it may play. That is not anyone choosing to make it worse; it is what happens when a gain instruction and a signal both act.

The more awkward part: normalisation is unpredictable

The third example is the key: a live take's dynamics are wholly unpredictable — very loud this time, very quiet the next.

The algorithm must analyse the file first, and what it analyses is a single statistical value. The result:

• a live recording that is mostly quiet with one climax → that climax gets flattened;

• a quiet passage followed by an explosion → the decision looks arbitrary.

This is why live versions often sound "wrong" dynamically on streaming. It is not a performance problem. It is a statistics problem.

A verifiable phenomenon

Take three records from different decades, play them on a streaming service, then:

1. compare the same piece across recordings, with all playback effects off;

2. notice the relative differences between quiet passages;

3. then notice where the loudest moment sits.

You will find the overall level is now similar, but the shape of the dynamics has changed. Normalisation adjusts "the whole", not "each moment".

What to do in practice

Three pragmatic steps:

1. Listen to originals — when you can find a physical record or a non-normalised file, listen to that;

2. Watch the dynamic shape — if a passage should have an explosion and you cannot hear it on streaming, the explosion may not be missing; it may be compressed out;

3. Never use loudness as a quality judgement — loud is not good. This matters more in the normalisation era, because loudness is no longer the musician's choice.

A position

Normalisation is itself a well-intentioned technical decision: it lets a 1900 recording be heard to the end on a phone.

The price is paid in dynamics — and dynamics are the most direct expression of "what is happening at this moment".

The first example is where that price is most visible: after normalisation, a climax becomes "a little louder" rather than a release.

So the real loss is not fidelity. It is the information contained in "rise and fall" itself.

Listen: what a flattened dynamic looks like

Below is first a quiet four-note motif, then a sudden note an octave higher. Imagine two playbacks: one preserving that jump (dynamics intact), one levelling the whole thing (the jump becomes almost inaudible).

Next

• What dynamics are: Dynamics.

• Where dynamics get processed: mixing and mastering.

• Its upstream layer: Timbre.

• Another answer to the same problem: the vinyl revival.

Sources

• 本文为响度归一原理的通识性说明；不涉及任何平台的具体规则数值与实施细节
• 涉及的曲目均标注来源与许可档位，可在线试听

Used for fact-checking only; all prose is original.