AI fingerprint cleanup, before and after: how to check a Suno track yourself

October 16, 2026 · All posts

People who search for an "AI audio fingerprint remover" or "remove AI artifacts from Suno" usually want one answer: what happens to my song when I run it through cleanup, and how do I know nothing was damaged? This article is a checklist for that. It uses the numbers from 2,230 cleaned tracks, and it shows how to run the same check on your own file.

What cleanup changes in the file

A generator leaves four kinds of marks in the audio. Cleanup handles each with one small, specific fix:

• narrow tones above 15 kHz are cut out with very narrow filters, so the gap is hard to hear;
• if such marks are found, the band above about 15.5 kHz is replaced with natural "air", so the track does not turn dull;
• small level jumps at the seams between generated blocks are smoothed, about every half second, while real drum hits and drops are detected and left alone;
• if the spectrum shows a row of fixed peaks, the pitch moves by 5.5 to 8.5 cents, less than a tenth of a semitone, and the length stays exact to the sample.

Notice what is missing from that list. There is no compression, no equalisation for loudness and no tempo change. Cleanup does not master the song. If a cleanup changed the loudness of a track a lot, it would be mastering it without saying so, and that is the first thing to check.

Before: measure the export

Measure the Suno export before you clean it. You need four readings: integrated loudness in LUFS, loudness range (LRA), true peak, and the crest factor. Stereo correlation is useful too for a track with a wide mix. The free LUFS checker gives loudness, true peak and loudness range. Crest factor and stereo correlation need a meter that shows them.

Write the four or five numbers down, or save them in a text file next to the export. Without a "before" you cannot tell what cleanup did, only what the track sounds like now.

After: what the numbers should look like

We cleaned 2,230 tracks between 23 July and 8 October 2026, excluding our own test runs. Each track was measured as uploaded and again after cleanup, with the same meters. The changes were small:

Output true peak was even clearer: all 1,765 measured tracks came out at or below −1 dBTP, with a median of −1.36 dBTP. The cleaned file will not clip when a streaming service encodes it.

Treat these as the range of a healthy cleanup, not as a promise about your song. Here is how to read your own numbers against it:

• Loudness within about ±0.5 LU and LRA within about ±0.5 LU is the normal result. Most tracks land here.
• A change beyond ±1 LU is not automatically a fault, but look at the level of the original first (see the next section). If the track was not especially loud and still moved that far, measure the file again with the same meter and settings.
• True peak above −1 dBTP after cleanup should not happen. Treat it as a fault: do not release that file, measure it again, and keep the original.
• Stereo correlation moving by more than 0.01 is rare. Check the cleaned file on headphones and on a speaker before you accept it.

Why a loud track can get quieter

For most tracks the loudness did not move at all. The tracks that lost more than 1 dB had one thing in common: they were already very loud. Tracks louder than −10 LUFS lost a median of 0.84 dB; everything else, 0.02 dB.

The reason is simple. After cleanup, peaks are kept under −1 dBTP. A track that was pushed to the ceiling has nowhere to go but slightly down. This is not damage, and it is not a reason to skip cleanup. It is a reason to master the cleaned file afterwards, to a level your platform expects. Spotify plays music at about −14 LUFS, so a track that is louder than that gets turned down anyway.

A before-and-after check on your own track

This takes about ten minutes for one song:

1. Export the track as WAV, not MP3. Compressed files hide the marks that cleanup looks for, and they also make the comparison less fair.
2. Measure the export and write the numbers down.
3. Clean it in the lab: choose "Clean AI artifacts". Cleanup never spends a credit.
4. Measure the cleaned file with the same meter and the same settings.
5. Compare the two numbers against the table above.
6. Listen to both files on the same speakers, at matched loudness. This step matters most. A difference of 0.3 dB is enough to make the quieter file sound duller, even when nothing else has changed. If you cannot match the level, compare the two files only for the sound of the highs, and do not judge the loudness.
7. Only after this, master the cleaned file for your platform. The guide to mastering a Suno song explains the steps and the targets.

If the cleaned file passes the numbers and you prefer the original when you listen, keep the original. Cleanup is an option, not a requirement.

What before and after cannot show

Be clear about the limits, because they matter as much as the numbers:

• Where the melody, words and timbre come from. Cleanup does not change how the model made the song. The metallic sheen and the watery highs that come from the model itself stay. A better prompt or a new generation fixes those; post-processing does not.
• Which generator made the track. Our numbers describe the cleanup, not one tool. People clean songs from several generators, and file names rarely say which one made them.
• How a detector will judge the track. We do not promise any result for that. Neural detectors that judge timbre and structure still recognise generated music after any processing, ours included. If a platform or distributor asks you to disclose AI use, disclose it.

Where to read more

The step-by-step table for each cleanup stage is on the AI artifact cleanup page. For the method behind the 2,230-track measurement, read What AI artifact cleanup actually changes. To check loudness and true peak before you start, use the LUFS checker.

Only aggregate numbers are published here: no audio, no titles and no user data. You may cite them with a link to this page.