What AI artifact cleanup actually changes: 2,230 tracks measured

October 8, 2026 · All posts

Search for "AI fingerprint remover" or "remove AI artifacts from music" and you will find plenty of promises and very few numbers. A cleanup step should fix one thing without breaking the rest of the track, so we checked exactly that: what our cleanup did to the sound of every track people ran through it.

What cleanup does

Cleanup looks for the marks a generator leaves in the audio and removes them (the full step-by-step table is on the AI artifact cleanup page):

• narrow tones above 15 kHz are cut out with very narrow filters;
• the top band above about 15.5 kHz is replaced with natural "air", only when such marks were found;
• small level jumps at the seams between generated blocks are smoothed, while real drum hits are left alone;
• when the spectrum shows a row of fixed peaks, pitch shifts by 5.5–8.5 cents — less than a tenth of a semitone — and the length stays exact.

It does not master the track: loudness and tone are meant to stay where they were. The question is whether they do.

The numbers

We took every cleanup completed between 23 July and 8 October 2026, excluding our own test runs: 2,230 tracks. Each track was measured as uploaded and again after cleanup, with the same meters. Not every early record has every reading, so each row shows how many tracks it is based on.

And the output true peak: all 1,765 measured tracks came out at or below −1 dBTP (median −1.36 dBTP), so the cleaned file will not clip when a streaming service encodes it.

Why some tracks got a little quieter

For most tracks loudness did not move at all. The 5% that lost more than 1 dB have 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.

That is not the cleanup itself. After cleaning, peaks are brought under −1 dBTP, and a track that was pushed to the ceiling has nowhere to go but slightly down. Streaming services would turn it down by more than that anyway — Spotify plays music at about −14 LUFS.

What these numbers do not show

• How often the marks were found. Until today the cleanup kept its findings — marks cut, pitch shift applied — only for the person who ran it. From 8 October they are recorded with every job, and we will add the share of tracks with marks here once there is enough data.
• Which generator made each track. People clean songs from Suno and other tools, and file names rarely say which. These numbers describe the cleanup, not one generator.
• How it sounds to you. Numbers show that the track was not damaged. Whether the result is better is a listening question: compare the original and the cleaned file on the same speakers.

What cleanup cannot change

The metallic sheen and watery highs that come from the model itself stay — a better prompt or a re-generation fixes those, not post-processing. Cleanup also does not change where the track came from. Neural AI detectors judge timbre and structure and still recognise generated music after any processing, ours included. If a platform or distributor asks you to disclose AI use, disclose it.

How to use it

1. Measure the export with the free LUFS checker.
2. Clean it in the lab — choose "Clean AI artifacts". Cleanup never spends a credit.
3. Master the cleaned file for your platform: see how to master a Suno song.

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