You generated a track on Suno, it sounded great in the browser, and then you dropped it into a DAW next to real recordings — and something is off. A metallic sheen on the cymbals. Sustained pads that swirl like they are underwater. Vocals that smear on long notes. Nothing you do with EQ quite removes it.
Those are codec artifacts, and they are a property of how Suno makes audio rather than a flaw in your particular generation. This guide covers what each one is, the export setting that quietly makes them worse, and how to actually remove them. For the same treatment across Udio, Stable Audio and MusicGen, see the general artifact guide.
Short on time? de-artifact removes these automatically. Clean a track free in your browser — no install — or run the plug-in on your master bus with a 14-day full trial.
Suno does not generate a waveform sample by sample. A language model generates a sequence of discrete tokens, and a neural audio codec decodes those tokens into audio. The codec compresses sound using residual vector quantization (RVQ): it approximates the signal with a first codebook, measures what is left over, approximates that residual with a second codebook, and so on.
Each stage leaves a little error behind. What survives into the decoded audio is not random noise — it is structured, with a consistent spectral signature. That is precisely why a model can be trained to recognise and subtract it, and why a graphic EQ cannot: the residual lives in the same frequency range as the cymbals and air you want to keep.
Suno's free Basic plan exports MP3. WAV is available on the Pro and Premier tiers. That distinction matters more than it looks:
The trap paying users fall into: per Suno's own documentation, downloading from mobile defaults to MP3 even on Pro and Premier. If you are paying for WAV, download from suno.com on a desktop browser — otherwise you are cleaning a lossy file without realising it.
Worth knowing: exporting as WAV does not remove artifacts — it stops you from adding new ones. If you plan to clean a track, start from the highest-quality export you can get; every removal step works better on one layer of damage than on two.
| What you hear | Where it lives | Cause |
|---|---|---|
| Metallic shimmer, artificial brightness | Cymbals, hats, vocal sibilance | High-frequency RVQ quantization residual |
| Watery, swirling movement | Pads, strings, held vocal notes | Phase inconsistency in the codec's reconstruction |
| Grainy, harsh texture above ~8 kHz | Air band, room tails | Bandwidth limits in the codec's learned spectral model |
| Thin, hollow mids | Lead vocal, snare body | Token-domain dynamics translating poorly to waveform dynamics |
| Pre-echo, smeared transients | Drum hits, plucked strings | MP3 export only — added by the lossy encoder |
de-artifact is a VST3 / AU / CLAP plug-in built on ArtifactNet, a forensic network trained to isolate the codec residual and subtract it while leaving the music. It runs on the master bus in real time, so you can hear the result while you adjust it.
WAV if your plan allows it. If you are on Basic and only have MP3, that is fine — just pick the matching preset below.
After any compression or limiting. It should see the final signal, the same one a listener would hear.
MP3 export → Conservative, tuned for lossy sources so it does
not reinforce MP3 ringing. WAV export → Default, moving to
Aggressive if shimmer survives. Watery sustains or smeared vocals
→ More Harmonics. Isolated vocal stems →
Vocal.
Diff solos what is being removed. You want to hear shimmer, hiss and swirl. If you can make out melody, drums or words, you are subtracting music — back off the strength until you cannot.
Clean one in the browser with no install, or take the plug-in for 14 days.
That is a separate question with a separate answer — the same residual that sounds wrong is also what detectors key on. If that is what brought you here, see how AI-generated music is detected.
Suno does not synthesise a waveform directly. A language model produces discrete tokens, and a neural audio codec decodes those tokens back into audio. That codec uses residual vector quantization, and the quantization error survives decoding as a structured high-frequency residual — heard as metallic shimmer or an artificial sheen. It is a property of the generation pipeline, not of your mix.
No. WAV avoids adding a second, lossy MP3 layer, which is worth doing — but the codec residual is baked in before any file is exported, so it is present in the WAV too. WAV gives you a cleaner starting point for removal, not a clean track.
They can mask them. The residual overlaps the same frequency range as cymbals, air and vocal consonants, so a filter wide enough to suppress it also dulls the music. Removal has to key on the artifact's spectral structure rather than its frequency band.
Start from the source format. For an MP3 export use Conservative, which is tuned for lossy-encoded sources and applies lighter subtraction so it does not reinforce MP3 ringing. For a WAV export start at Default and move to Aggressive if the shimmer persists. If the problem is watery sustains or smeared vocals rather than brightness, use More Harmonics.
Suno's downloads default to MP3 on mobile regardless of plan. WAV requires downloading from suno.com in a desktop browser. It is worth checking the file you actually have before assuming you are working from a lossless source.
Yes. The browser demo cleans a track without an install, and the plug-in has a 14-day full trial with every preset and parameter unlocked.