AI Music Artifact Removers Compared: 8 Tools Tested on Suno and ACE-Step Tracks

If your Suno track sounds metallic, watery or smeared, several AI music artifact removers now promise to fix it. We ran eight of them, from free tools like Shimmer, Deshimmer and InstaSong Desunofier to plug-ins like soothe2, RX, UNCHIRP and de-artifact, on the same recordings with frozen settings. We matched the loudness and published the audio, spectrograms and measurements so you can compare them yourself.

Disclosure: we make de-artifact, one of the tools tested, so we have a commercial interest in the outcome. The method, every measurement and lossless copies of every output are linked below.

How do AI music artifact removers change a track? Listen and compare

Choose one of the 16 shared examples, then switch between versions. The audio and its spectrogram change together, and switching keeps your playback position. Every version of an example covers the same aligned interval at the same loudness (normally −20 LUFS), with no limiter.

Choose an examplesuno_01 · Suno · ai mixsuno_02 · Suno · ai mixsuno_03 · Suno · ai mixsuno_04 · Suno · ai mixace_01 · ACE-Step · ai mixace_02 · ACE-Step · ai mixace_03 · ACE-Step · ai mixace_04 · ACE-Step · ai mixstem_01 · Suno separation · stemstem_02 · Suno separation · stemstem_03 · Suno separation · stemstem_04 · Suno separation · stemhuman_01 · Open Goldberg Variations · clean humanhuman_02 · Open Goldberg Variations · clean humancodec_01 · Open Goldberg Variations · codeccodec_02 · Open Goldberg Variations · codec

suno_01 · Suno · ai mix · playback target -20.00 LUFS

Originalde-artifact 0.4.7soothe2 (legacy)Gullfoss (neutral factory control)RX 10 Spectral De-noise (legacy)Acon DeNoise2Shimmer 2.3.5DeshimmerBypass−3 dB high-shelf controlInstaSong Desunofier (Strong)UNCHIRP 1.2.0 (factory)

Selected: suno_01 / Original

Download this version as lossless FLAC

The player streams Opus at 256 kbps. All measurements and spectrograms were made on the 24-bit PCM files, and each version can be downloaded as lossless FLAC. Colour differences in a spectrogram show where energy changed, not how much of an artifact was removed. stem_04 is a near-silent stem kept because the input set was frozen before processing; it is not a useful musical example.

Which AI music artifact removers did we test?

Eight configurations: de-artifact, soothe2, RX 10 Spectral De-noise, Acon DeNoise2, Shimmer 2.3.5, Deshimmer, InstaSong Desunofier and UNCHIRP. Three controls sit alongside them: bypass, a fixed −3 dB high-shelf EQ, and Gullfoss at factory settings (Recover and Tame at zero, so it is a neutral workflow check, not a test of Gullfoss). We used installed licences and free tools. RX 10 and soothe2 are older versions than the current RX and soothe3, and the de-artifact we measured is 0.4.7; the current release is 0.4.8.

Settings were fixed on eight separate calibration tracks before any evaluation file was processed: documented restoration guidance where it exists, factory defaults otherwise, mastering off wherever a tool exposes it. No setting was tuned on evaluation outputs. New spending for the study was 0 KRW.

How much does each artifact remover change the audio?

The change figure is the residual after aligning each output to its input and fitting one overall gain, relative to the input, in dB. It answers "how much did this tool alter the recording?" It does not say whether the alteration removed an artifact or improved the music. The fixed high-shelf EQ shows what a simple tonal change scores.

Common 16

Codec damage with a known original

Two piano recordings were damaged with fixed 64 kbps MP3 encoding, so each has a lossless original to compare against. Error is RMS error relative to the original, without a gain fit; LSD is log-spectral distance. Lower means closer to the original, for this codec and these two files only.

Tool by tool: what each remover did to the same tracks

The same figures as the tables above, one tool at a time. Gullfoss, the high-shelf EQ and bypass are controls and are left out here. A larger change means the tool did more to the recording; whether that removed artifacts or added colour is something to judge by ear in the player.

Acon DeNoise2

Local plugin. Listed price on October 9, 2026: Restoration Suite 2.x USD99, VAT included where applicable; existing DeNoise2 measured. On the common 16 it changed the audio by a median of −14.06 dB (range −23.43 to −10.15 dB) and moved loudness by −1.90 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 16.20 dB. Median band changes, low to high: 20–200 Hz +0.04 dB, 200 Hz–2 kHz −0.35 dB, 2–5 kHz −1.38 dB, 5–10 kHz −1.08 dB, 10–20 kHz −3.05 dB.

de-artifact 0.4.7

Local plug-in (VST3, AU, CLAP) and a separate Cloud version. Listed price on October 9, 2026: USD 149 one-time release price (USD 179 list), 14-day trial. On the common 16 it changed the audio by a median of −27.61 dB (range −109.83 to −20.35 dB) and moved loudness by −0.30 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 15.46 dB. Median band changes, low to high: 20–200 Hz −0.04 dB, 200 Hz–2 kHz +0.01 dB, 2–5 kHz −0.66 dB, 5–10 kHz −0.82 dB, 10–20 kHz −0.20 dB.

Deshimmer

Local Python DSP. Listed price on October 9, 2026: Free (open source). On the common 16 it changed the audio by a median of −25.78 dB (range −29.31 to −9.95 dB) and moved loudness by −1.10 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 15.32 dB. Median band changes, low to high: 20–200 Hz +0.03 dB, 200 Hz–2 kHz −0.15 dB, 2–5 kHz +0.49 dB, 5–10 kHz +0.69 dB, 10–20 kHz +0.57 dB.

InstaSong Desunofier (Strong)

Local browser DSP; no server upload per official page. Listed price on October 9, 2026: Free. On the common 16 it changed the audio by a median of −20.81 dB (range −36.41 to −0.25 dB) and moved loudness by −0.10 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 11.96 dB. Median band changes, low to high: 20–200 Hz +0.01 dB, 200 Hz–2 kHz −0.00 dB, 2–5 kHz −0.60 dB, 5–10 kHz −6.04 dB, 10–20 kHz −4.42 dB.

RX 10 Spectral De-noise (legacy)

Local plugin/standalone. Listed price on October 9, 2026: Existing RX10 installation measured; current RX12 price not verified from dynamic official store. On the common 16 it changed the audio by a median of −9.85 dB (range −19.82 to −3.40 dB) and moved loudness by −2.20 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 15.45 dB. Median band changes, low to high: 20–200 Hz −0.65 dB, 200 Hz–2 kHz −0.37 dB, 2–5 kHz +1.28 dB, 5–10 kHz +1.48 dB, 10–20 kHz +0.96 dB.

Shimmer 2.3.5

Local Python. Listed price on October 9, 2026: Free on owned/controlled computers, including commercial work; source-available restrictions. On the common 16 it changed the audio by a median of −44.68 dB (range −87.10 to −30.84 dB) and moved loudness by +0.00 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 15.39 dB. Median band changes, low to high: 20–200 Hz +0.00 dB, 200 Hz–2 kHz +0.00 dB, 2–5 kHz −0.01 dB, 5–10 kHz −0.02 dB, 10–20 kHz −0.01 dB.

soothe2 (legacy)

Local plugin. Listed price on October 9, 2026: Existing installed soothe2; current soothe3 20-day trial. On the common 16 it changed the audio by a median of −23.38 dB (range −76.99 to −16.92 dB) and moved loudness by −0.50 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 15.32 dB. Median band changes, low to high: 20–200 Hz +0.21 dB, 200 Hz–2 kHz −0.07 dB, 2–5 kHz −0.47 dB, 5–10 kHz −0.12 dB, 10–20 kHz +0.04 dB.

UNCHIRP 1.2.0 (factory)

Local plugin. Listed price on October 9, 2026: USD 269 / EUR 279 incl. 19% VAT; 30-day trial. On the common 16 it changed the audio by a median of −19.00 dB (range −23.67 to −6.43 dB) and moved loudness by +0.10 LU. On the two codec-damaged files, log-spectral distance to the original went from 15.32 to 14.23 dB. Median band changes, low to high: 20–200 Hz −0.06 dB, 200 Hz–2 kHz −0.02 dB, 2–5 kHz −0.65 dB, 5–10 kHz −0.89 dB, 10–20 kHz +2.42 dB.

How was the comparison run?

The evaluation set has 48 distinct sources: 16 AI full mixes (8 Suno, 8 ACE-Step), 16 separated stems, 8 codec-damaged recordings and 8 clean human recordings. The common 16 used in the player is 8 AI mixes (4 Suno, 4 ACE-Step), 4 stems, 2 codec files and 2 clean recordings. Every input is 30 seconds of 44.1 kHz stereo PCM24.

Suno files come from our own paid account; ACE-Step tracks were generated from text under the model card's commercial-output terms. Clean and codec music is Kimiko Ishizaka's CC0 Open Goldberg Variations, so the human material is solo piano and does not cover other genres. Some Suno sources were lossy exports, and converting them to WAV does not restore what was lost.

The analyser was checked first against known cases: bypass, a known gain, a 137-sample delay, a known EQ, clipping and a near-silent input. It measures alignment, level, loudness, true peak, clipping, silence, stereo correlation, band energy and time-resolved change. Listening copies share one integer-aligned interval and one scalar gain per file; if any version would exceed −1 dBTP, the whole example set uses a lower target instead of a limiter. Spectrograms use a 4,096-sample Hann STFT with a 1,024-sample hop on those copies.

What this study does not tell you: which tool sounds best, what share of artifacts any tool removed, or whether listeners prefer one result. Those need a listening test, which we did not run.

Data and downloads

• Per-file measurements (CSV) and full measurements with repeats (JSON)
• Source, rights and hash manifest
• Listening-copy crop, gain and hashes and spectrogram method and hashes
• Product survey
• Methods package: analyser code, frozen settings and data

Vendor software, model weights and licence details are not redistributed; reproducing a tool's output needs your own installation and the frozen settings in the package.

Rights references: Suno commercial use, ACE-Step model card, Open Goldberg Variations.

Related reading

• What is an audio artifact? The kinds of damage these tools try to remove.
• Suno metal artifacts: fizzy cymbals and guitars, and how to fix them.
• WAV vs MP3 for AI music, and why a lossy export limits any cleanup.
• How to clean a separated vocal stem.
• de-artifact walkthrough: what to listen for, chapter by chapter.
• de-artifact vs artefactFX: two different goals.

Want to try an AI music artifact remover on your own Suno track? de-artifact has a 14-day trial.

FAQ

What is the best AI music artifact remover?

This comparison does not name one. We measured how much eight tools changed the same Suno, ACE-Step and human recordings and how close codec-damaged files came to their originals, but we did not run a listening test. Use the matched-loudness player to judge by ear on the examples closest to your material.

How were the tools set up?

Settings were fixed on eight separate calibration tracks before any evaluation file was processed: documented restoration guidance where a vendor gives it, factory defaults otherwise, and mastering off wherever a tool exposes it. No setting was tuned on the evaluation outputs.

Why is the player audio Opus instead of WAV?

Streaming 194 lossless files would be slow, so the player uses Opus at 256 kbps. All measurements and spectrograms were made on the 24-bit PCM files, and every version can be downloaded as lossless FLAC from the player.

Is there a free Suno shimmer remover?

Yes. Of the tools we measured, Shimmer, Deshimmer and InstaSong Desunofier were free to use as of October 9, 2026, and all three ran on the full common set. Shimmer and Deshimmer run locally in Python; Desunofier runs in the browser. Their measured changes are in the tool-by-tool section.

How do I remove artifacts from a Suno song?

Run the track through an artifact remover, then compare the result with the original at matched loudness. Start from the best-quality export you have, because a lossy MP3 has already lost information no tool can restore. The examples in this comparison let you hear what each tool does before you commit to one.

Does an artifact remover change the rest of the mix?

Every tool changed clean human piano recordings too, by a median of about 9 to 37 dB below the input depending on the tool. Small changes on clean material suggest a tool leaves good audio alone. A large change is a reason to listen closely, not proof of damage.

Why are some tools missing?

We measured only tools we could run on the full common set with an export we were allowed to use. Products limited by watermarked trials, free credits that did not cover the set, or no macOS build are listed in the survey as not measured, which says nothing about how they perform.