If you've generated music with Suno, Udio, Stable Audio, or MusicGen, you've probably noticed it: a subtle metallic sheen, a watery swirl in the high mids, or a kind of hollow compression that makes the track feel synthetic even at first listen. These are AI music artifacts — and they come from the neural audio codec at the heart of every modern music generator.
This guide explains what they are, why they happen, and the most effective ways to remove them.
Modern AI music generators — Suno, Udio, Stable Audio, MusicGen, Riffusion — don't generate audio directly. They generate a sequence of discrete tokens using a language model, then decode those tokens into audio using a neural audio codec such as EnCodec or DAC.
These codecs use a technique called Residual Vector Quantization (RVQ) to compress audio into tokens. RVQ introduces quantization residuals — errors that survive into the decoded audio as repeatable, structured artifacts. They're not random noise: they have a specific spectral signature that trained ears (and trained models) can identify.
| Artifact type | What it sounds like | Root cause |
|---|---|---|
| RVQ ghosting | Metallic shimmer, artificial brightness, "digital" sheen | High-frequency codec residuals from RVQ quantization error |
| Codec residue | Watery, swirly modulation on sustained notes and pads | Phase inconsistencies in the codec's overlap-add reconstruction |
| HF aliasing | Harsh, grainy texture above 8–12 kHz | Bandwidth limitation artifacts from the codec's learned spectral model |
| Hollow compression | Mid-range sounds thin, lacks body and presence | Token-domain dynamics that don't translate naturally to waveform dynamics |
Key point: AI music artifacts are not mastering problems. EQ and compression can mask them temporarily, but they're baked into the audio at the codec level. Effective removal requires understanding the artifact's spectral structure.
de-artifact is a VST3 / AU / CLAP plug-in powered by ArtifactNet, a forensic neural network trained specifically to isolate and subtract RVQ codec residuals from AI-generated music. It's designed to sit on the master bus (2-bus) and process the full mix in real time.
Drag the MP3 or WAV file onto a new audio track. Route it to your master bus. Any DAW on macOS or Windows works — Logic Pro, Ableton, Reaper, Pro Tools, Cubase, Studio One.
Place de-artifact after any existing master bus processing. If your chain is: compressor → limiter → de-artifact, that's correct. de-artifact should see the final, loudness-normalized signal.
Start with AI Aggressive for Suno or Udio tracks. For Stable Audio or MusicGen, try HPSS Harmonic Focus — it preserves more high-frequency energy while targeting the codec residual.
The Aggressiveness knob controls how strongly the model subtracts the detected artifact. For most AI tracks, 50–70% gives the best balance. If the result sounds dull or over-processed, pull back to 30–40%.
Toggle Diff mode to solo the subtracted artifact signal. You should hear spectral noise and shimmer — not musical content. If you hear melody or drums, reduce Aggressiveness.
14-day full trial — every preset, every parameter. No card required.
If you don't want a dedicated plug-in, you can reduce (not eliminate) AI music artifacts using standard DAW tools. These approaches mask the artifact rather than removing its source.
RVQ ghosting concentrates between 6–14 kHz. A dynamic EQ set to reduce gain by 2–4 dB when energy exceeds a threshold in that range can smooth the metallic sheen without dulling the top end. This works best on tracks where the artifact is mild.
Codec residuals tend to be wider in the stereo field than the underlying music. Applying mild saturation (0.5–1 dB drive) to the side channel only can add harmonic density that masks the artificiality. Use a mid-side EQ to roll off above 12 kHz on the side channel afterward.
Manual spectral repair changes the tonal character of the track. You're trading one kind of artificiality (RVQ shimmer) for another (over-processed top end). For production-ready results — distribution, sync licensing, client delivery — a model-based approach like de-artifact gives cleaner output because it subtracts the artifact specifically, not the frequency band it inhabits.
| Generator | Recommended preset | Notes |
|---|---|---|
| Suno v3/v4 | AI Aggressive | Suno uses a high-bitrate codec; aggressive setting handles the HF residual well |
| Udio | AI Aggressive or HPSS Harmonic Focus | Udio tracks often have more harmonic content in the artifact — harmonic focus mode is gentler |
| Stable Audio | HPSS Harmonic Focus | Stable Audio's codec produces a different residual profile; harmonic focus preserves more top end |
| MusicGen / Riffusion | Lossy Conservative | These generators have a heavier codec imprint; conservative mode avoids over-subtraction |
| AI vocal stem (any) | AI Vocal Stem | Tuned for narrowband signals; avoids removing sibilance alongside the artifact |
If you're working with stems separated by Demucs, htdemucs, or Spleeter — rather than full AI-generated mixes — you may be dealing with a different type of artifact: stem bleed. This is when audio from one stem (e.g., drums) leaks into another (e.g., vocals) because the separator's ML model didn't fully isolate the sources.
Stem bleed is not the same as RVQ codec residuals, and de-artifact is not designed to fix it. de-leak-rt is a separate plug-in built specifically for this: a real-time VST3 / AU / CLAP leak gate with an ML classifier (F1 = 0.993) that detects inter-stem bleeding per-frame and applies multiband gain reduction only where leakage is detected. It works on vocal stems, instrument stems, and any Demucs / htdemucs / Spleeter output.
Yes — and it works by analyzing the same codec artifacts described above. ArtifactNet, the forensic model powering de-artifact, achieves F1 = 0.9829 on a held-out set of 6,183 tracks across 22 AI generators. It identifies AI-generated music by its residual fingerprint, not by generator-specific patterns — which means it generalizes to generators it hasn't seen during training.
The free online demo lets you upload any track and get a forensic verdict — AI-generated or human — along with a breakdown of the detected artifact signature. No account required.
de-artifact reports 500 ms latency in Standard mode for host PDC (Plugin Delay Compensation). During a DAW freeze or offline bounce, some hosts signal "offline mode" mid-stream, causing the plug-in to re-prime its processing buffer. The output is correct from the first audible frame onward. This is a one-time priming artifact at the very start of the bounce, not a gap during playback.
Lower Aggressiveness (try 30–40%), or switch to HPSS Harmonic Focus mode, which targets the percussive split of the artifact and preserves more high-frequency bandwidth. Pulling the high-shelf emphasis parameter to −3 dB also tells the model to be gentler on the top end.
Yes — the Lossy Conservative preset is specifically tuned for lossy-encoded sources where MP3 compression artifacts are layered on top of the codec residual. It applies lighter subtraction to avoid reinforcing the MP3 ringing.
de-artifact is trained to target the specific spectral signature of RVQ codec residuals. On real music without AI codec artifacts, Aggressiveness below 30% has minimal effect. Above 50% on a clean real-music source you may hear subtle tonal thinning — keep it below 20% if you're processing mixed sources.