How to Compare Audio Before and After Loudness Normalisation

Normalisation is not just “make it louder”

Every platform has a loudness target. Streaming services, podcast directories and broadcast standards all want your audio at a specific integrated level, and something will normalise it whether you do it yourself or not.

If you only turn the whole file up or down, nothing else changes — that’s gain, and it’s harmless. But most normalisation workflows do more: they apply limiting to stop peaks clipping, sometimes compression to raise the average level, and occasionally multiband processing that changes the tonal balance.

Those extra stages are where audio quietly gets worse, and a loudness meter won’t tell you. It reports that you hit the target. It says nothing about what you gave up to get there.

What to check after a normalisation pass

Did the dynamics survive? Heavy limiting flattens the difference between loud and quiet passages. On music this shows up as a loss of impact; on speech it makes everything sound the same weight. If your quiet passages got proportionally louder than your loud ones, the dynamic range shrank.

Did anything clip? A limiter can be pushed past the point where it stops being transparent. Short bursts of distortion on transients — drum hits, plosives, door slams — are the classic symptom, and they are easy to miss on small speakers.

Did the tone change? Multiband processing can shift the balance without you asking it to. Boomier low end or a harsher top end after normalising means the process touched more than the level.

Did the noise floor come up? Raising the average level raises everything, including hiss, room tone and hum. Audio that sounded clean at the original level can reveal a noise floor you did not know was there.

Comparing the two files

Upload the original and the normalised version to DiffALL’s audio comparison. You get an overall similarity score, a per-second chart, and mel spectrograms of both files plus a difference view.

Because a straight gain change barely moves the analysis, the score tells you something specific: how much of the difference is level, and how much is processing.

• 95%+ — essentially a level change. The processing was transparent.
• 85–95% — real processing happened but stayed reasonable. Worth listening to the flagged sections.
• Below 85% — the normalisation did substantial work. On a mastering chain that may be intended; on a “just hit the target” pass it is not.

Reading the per-second chart

This is where a loudness meter has nothing to offer. The chart shows similarity second by second, so you can see when the processing worked hardest.

A flat line near the top means the processing was even across the file. Good sign.

Dips at specific moments mean the limiter engaged hard there. Check those seconds. On speech they are usually plosives or a raised voice; on music they are transients. If a dip lines up with a moment that matters musically, you have found your problem.

A sagging section means a whole passage got squashed — typically the loudest part of the file, where a compressor sat on it for several seconds. This is what pumping looks like before you can hear it.

Comparing the spectrograms

The mel spectrograms show frequency content over time for both files, plus a difference map. This is the fastest way to answer “did the tone change”.

A difference map that is evenly quiet means the processing was level-only. Bright horizontal bands mean specific frequency ranges moved — a band at the bottom is the low end getting pushed, a band at the top is the high end. Either one means your normalisation did EQ you may not have intended.

A workflow that catches problems early

1. Keep the pre-normalisation master. You cannot compare against something you overwrote.
2. Normalise as a separate step, not baked into your export. If you cannot produce a before-and-after pair, you cannot check the work.
3. Compare before you publish. Two minutes of checking beats discovering distortion after a thousand people have downloaded it.
4. Compare again after any platform change. If a service changes its target, your existing chain is now normalising to the wrong number.

When a low score is fine

A mastering pass is supposed to change the audio, sometimes a lot. A 70% similarity between an unmastered mix and a finished master is not a problem — it is the work.

The score is only meaningful against your intent. Use it to confirm that a step you expected to be transparent actually was, and to see exactly what a step you expected to be heavy-handed actually did.

Try it

Compare two audio files → — upload the original and the normalised version and get a similarity score, a per-second chart, and spectrograms of both.