You exported a Suno vocal, or pulled a vocal out of a finished mix with a stem separator, and something sits on top of the voice: a glassy, phasey, slightly metallic sparkle that people call shimmer. It is most obvious on held vowels and in the top end, and it makes the singer sound as if they are behind a sheet of glass. Deshimmer is also the name of a free open-source tool. This guide covers the problem itself and the tools for it: what shimmer usually is, how to tell it apart from three things that sound similar, and a chain that reduces it without taking the voice with it.
We make de-artifact, one of the tools mentioned below. We say where it fits, where it does not, and which free or general tools cover the neighbouring problems.
The short answer
Start from the best file you can get, confirm it is shimmer and not reverb, sibilance or a resonance, then reduce it with a residue or artifact remover and listen to what that tool removes. On a separated vocal, remove bleed first. Use a de-esser or dynamic EQ only for sibilance and harshness, and compare every step at matched loudness. Detail the generator or separator never kept cannot be brought back, so aim for less shimmer, not a perfect voice.
What “shimmer” means here
The word has three meanings in audio, and search results mix them up:
- The artifact this page is about: a glassy or metallic layer that follows a generated or separated vocal and stops when the voice stops.
- Shimmer reverb, an effect people add on purpose. Valhalla DSP’s ValhallaShimmer notes say the plug-in can “pitch shift the feedback signal” of the reverb, and that a +12 semitone shift with feedback at 0.5 or more gives the classic “shimmer” sound. If a shimmer reverb is on your vocal bus, switch it off. That is the whole fix.
- Shimmer in voice analysis, a measurement. The Praat manual defines local shimmer as “the average absolute difference between the amplitudes of consecutive periods, divided by the average amplitude.” That is a measure of the voice itself, not a processing artifact.
Where the artifact comes from
There is no single published cause. Three sources are plausible, and a vocal can carry more than one:
- Generation. We could not find a published architecture for the current Suno or Udio music models, so we do not claim how they produce audio. As an example of a related technique, Meta’s EnCodec neural codec encodes audio into a quantized latent space and decodes it back; its README describes residual vector quantization, with the number of codebooks set by the target bitrate. The EnCodec paper itself treats artifacts in the decoded audio as something to reduce in training. Whether that applies to your file is a hypothesis, not a diagnosis.
- Lossy export. A low-bitrate MP3, AAC or Opus file adds its own high-frequency hash. Our de-artifact manual describes it as a related class of artifact, since both are reconstructed from a compressed code.
- Separation. Source-separation research scores separation errors along separate axes: the mir_eval documentation, following Vincent, Gribonval and Févotte (2006), assesses an estimate in terms of the filtered true source, interference and artifacts. Shimmer on a separated vocal would fall under artifacts, and bleed under interference. They need different fixes.
Shimmer, reverb, sibilance or resonance?
Name the problem before you reach for a tool. Loop a phrase that has a held vowel, a consonant and a gap, solo the vocal and listen.
If you only have a full Suno mix, the music keeps playing in the gaps, so you cannot solo the vocal or run the gap test. You have two options. Check whether you can download stems: Suno’s help centre says individual stem files are available when stems exist for a song. Or skip the gap test, go to steps 3 and 4 below with the Default preset, and judge the result by what the tool removes (the Residual) rather than by the gaps.
| What you hear | Listening test | Likely problem | First tool to try |
|---|---|---|---|
| Glassy, phasey sparkle on every sustained note | It starts and stops with the voice; the gaps between phrases are clean | Shimmer (processing residue) | A residue or artifact remover |
| A wash that continues after the note ends | It decays in the gap instead of stopping with the voice | Reverb, either in the generation or a shimmer reverb on the bus | Bypass any reverb you added; regenerate a drier take |
| Sharp, hissy bursts | Only on s, sh, t and ch sounds, not on vowels | Sibilance | A de-esser or one band of dynamic EQ |
| A ring at one pitch | The same frequency pokes out across different notes | Resonance | A dynamic resonance suppressor |
| Snare, bass or chords in the gaps | You can hear the arrangement where the vocal should be alone | Bleed from separation | A bleed remover, before anything else |
The gap test comes from our guide to cleaning a separated vocal stem: residue is modulated by the voice and disappears in the gaps, while bleed does not.
Step by step: a shimmer-reduction chain
- Get the best source file. Suno’s help centre says MP3 downloads are available on all plans and WAV downloads to Pro and Premier subscribers on the website, with mobile downloads defaulting to MP3. If you can export WAV, start there. Converting an MP3 to WAV does not restore what the MP3 discarded; see WAV vs MP3 for AI music.
- Remove what is not shimmer. Switch off any shimmer reverb or chorus you added. On a separated stem, deal with bleed first. Our vocal stem guide puts bleed before residue, because a residue tool should not have to decide whether a snare hit is an artifact.
- Try gentle broadband approaches first. A high-shelf cut of a dB or two, or a regenerated take, is worth trying first on a mild case. A steep low-pass removes shimmer and air together, so check the voice still sounds open.
- Reduce the residue. Use a tool that estimates the artifact rather than a fixed frequency band. In de-artifact, the manual’s recipe for a separated vocal is: load the Vocal preset (described as harmonic-weighted, transient- and sibilance-aware), switch on Residual to hear only what is being removed, lower Strength if breaths and consonants disappear into it, and nudge Brightness up slightly if the result sounds dark. Strength runs from 0 to 200% with a default of 60%. On a full Suno or Udio mix, the manual suggests the Default preset first. Free options exist too. Shimmer and Deshimmer both run locally in Python and were free to use when we checked on October 9, 2026. Our comparison of AI music artifact removers ran both, alongside de-artifact and others, on the same recordings. It measured how much each tool changed the audio and ran no listening panel, so it does not say which one sounds best. We make one of the tools in it. Our guide to AI artifact removal plugins and tools explains how the open-source deshimmer project works and where each kind of tool fits.
- Treat adjacent problems with the right tool. For sibilance, use a de-esser or a single dynamic EQ band on the s sounds. TDR Nova, which its maker describes as “a parallel dynamic equalizer”, has a free version for Windows, macOS and Linux. de-artifact no longer has a separate de-esser; its manual suggests raising Resonance Depth to about 0.4 and Harmonic toward 2.5 for ringing sibilance. For one ring that follows the notes, a resonance suppressor such as our ENZYME is the tool, not a residue remover.
- Listen to the removed signal. Whatever tool you use, solo what it takes out. You want fizz, hash and glassy sparkle. If you hear the words, the breath or the shape of the consonants, back off. The de-artifact manual calls this the whole technique: is anything musical leaking into the residual?
- Compare at matched loudness. Cutting energy makes a vocal quieter, and the louder version usually wins a quick A/B. Match levels, then listen in the full mix, not only solo.
- Print a copy and keep the original. Keep the untouched file so you can redo the chain later.
Where de-artifact fits, and where it does not
de-artifact reduces metallic shimmer and watery textures in generated tracks and separated stems. It runs as a VST3 or CLAP plug-in on macOS and Windows and as an Audio Unit on macOS, with AAX for Pro Tools on Apple Silicon Macs only (stereo inserts). Our manual says artifacts concentrate above roughly 3 kHz, which is why the model is conservative on clean acoustic material. That is our own description, not an independent finding. What it removes can still include musical detail, so the Residual check matters on every file.
It is the wrong tool for bleed (that is de-leak-rt), for one ringing resonance (ENZYME or another suppressor) and for reverb printed into the take.
What cannot be fixed
- Detail that was never kept. If the generator, codec or separator discarded part of the voice, cleanup cannot recover it. de-artifact’s positive Brightness regenerates the band above 8 kHz, which the manual describes as the model’s estimate, not the original restored.
- Heavy cases without trade-offs. Past a point, reducing shimmer also thins the voice. When the Residual starts sounding like the singer, a new generation or a different separation may be the better fix.
- The performance itself. Pitch, phrasing and timing are untouched. A cleaned generation still sounds like the generation it came from.
- The question of origin. Reducing shimmer does not change whether a track is AI-generated. It is not a way to pass an AI detector, and we do not offer it as one.
Hear it on your own vocal first
Upload a file to the free browser preview: no account, a 30-second preview of the start of the file, WAV, MP3, FLAC, OGG, Opus, M4A or AAC up to 50 MB. It is a listening preview, not a full-track export. To work in your DAW, de-artifact has a 14-day trial; the plug-in is $149 one-time at the release price as of October 25, 2026.
After the preview, a free Intrect account comes with $1 of credit, about 4 minutes of browser cleanup, then cleanup costs $0.25 per minute of audio. Cloud credit and the plug-in are separate purchases.
Related guides
- How to clean a separated vocal stem
- Remove instrument bleed from a vocal stem
- Suno artifacts
- How to fix AI music artifacts
- Dynamic resonance suppressor plugins compared
- AI music artifact removers compared
- Remove AI artifacts: plugins and tools compared
- AI music restoration: what to run first
- What is an audio artifact?
Sources
- Valhalla DSP, ValhallaShimmer notes (PDF), valhalladsp.com/shimmer/ValhallaShimmerNotes.pdf, read October 25, 2026.
- Praat manual, Voice 3. Shimmer, fon.hum.uva.nl/praat/manual, read October 25, 2026.
- Défossez et al., High Fidelity Neural Audio Compression (EnCodec), arxiv.org/abs/2210.13438, abstract read October 25, 2026; EnCodec repository README, github.com/facebookresearch/encodec, read October 25, 2026.
- mir_eval separation module documentation, citing Vincent, Gribonval and Févotte, “Performance measurement in blind audio source separation”, IEEE TASLP 14(4), 2006, mir-eval.readthedocs.io, read October 25, 2026.
- Suno Help Center, What types of files can I download?, help.suno.com/en/articles/13926081, read October 25, 2026.
- Tokyo Dawn Records, TDR Nova, tokyodawn.net/tdr-nova, read October 25, 2026.
- Our de-artifact manual, product page and artifact remover benchmark (tool prices as listed October 9, 2026).
FAQ
What is shimmer on an AI vocal?
It is a glassy, phasey or metallic sparkle that rides on the voice and stops when the voice stops. It is not reverb, which keeps decaying after a note ends, and not sibilance, which sits only on s, sh and t sounds. On generated audio and separated stems it is usually a processing residue rather than part of the performance.
Can EQ or a de-esser remove shimmer?
Only partly. A static high cut dulls the whole voice along with the shimmer, and a de-esser acts only on sibilant consonants. Use them for those adjacent problems, and use a residue or artifact remover for shimmer that sits on every sustained note. In each case, listen to what the tool removes.
Does removing shimmer change whether a track counts as AI-generated?
No. Cleanup changes how a vocal sounds. It does not change who or what made the track, and it is not a way to pass an AI detector. Distributors set their own rules for AI-generated audio, so read them before you release.