Why AI Music Sounds Metallic — and How to Fix It
2026-09-069 min
Put a Suno or Udio export on headphones and you hear a hiss floor across the top end, plus a metallic ring around cymbals and sibilants. It is not a bug in one model. It falls out of how these systems generate audio, and there is not one cause but two. Separating them is what separates the fixes.
The short answer
- AI music artifacts come from two different mechanisms: a high-frequency noise floor left by residual vector quantization (RVQ), and stationary spectral lines created by the decoder's transposed convolutions.
- The dominant part of what you hear as hiss is the noise floor, not the lines. In Honenuki's measurements it concentrates around 8-12 kHz and is present even in passages where a human recording would be digitally silent.
- Broadband denoisers remove that floor along with cymbals and consonants. Notch filters can catch the lines, but most tools cannot place notches that high in the first place.
- Mastering over it makes it worse. Make-up gain lifts the noise floor along with the signal, and an air-band boost lands just above the band where the hiss is strongest.
- Honenuki treats the two on separate paths: a 6-14 kHz noise gate capped at -40 dB and a standing-line notch capped at -20 dB, combined with max rather than multiplied, so a hard pass on one path cannot compound the other.
1. Why AI-generated tracks hiss in the first place
Generative models do not write waveforms directly. They produce a compressed internal representation, and a neural codec decoder unfolds that into audio. Artifacts are left at both stages, with different characters.
Start with residual vector quantization. RVQ quantizes the internal representation in stages, and each stage takes whatever reduces the remaining error most. But musical energy sits overwhelmingly in the low end, so adding stages refines the low-frequency residual rather than reaching upward. Neural codecs have no per-band bit allocation the way MP3 does — the SoundStream paper states the rate budget is allocated uniformly across quantizers.
The training loss is also near-sighted up there. The mel spectrograms these codecs use grow coarser with frequency: at 44.1 kHz with 80 bands, one band spans roughly 74 Hz near 1 kHz but about 466 Hz at 10 kHz. Matching total energy in that band is enough to reduce the loss, so fine structure goes unconstrained.
The published failure mode follows. The DAC paper reports that generated spectrograms still look blurry, with over-smoothing artifacts in high frequencies. A decoder that is unsure about the top end does not leave it silent — it learns to fill it with a plausible noise texture. That texture is the hiss.
Now the transposed convolutions. The decoder upsamples its internal representation, and a transposed convolution does that by inserting zeros between samples. That zero-insertion periodizes the spectrum and produces regular peaks. The 2025 paper A Fourier Explanation of AI-music Artifacts proves this mathematically, and the important part is that the peaks are determined by architecture alone — not by training data or weights.
So these lines sit at absolute frequencies with no relation to the key or harmony of the track. Transpose the song and they stay put. That property is what decides the strategy: a target that does not move can be caught by a stationary filter.
The two mechanisms must not be conflated. RVQ error is signal-dependent and broadband, so it cannot produce evenly spaced lines; transposed convolutions produce lines, not a broadband floor. Treat them as one cause and you end up with one fix for both — which is exactly how top ends get destroyed.
2. How engineers remove it today
Existing tools get you a long way. They were not built for this problem, though, so each approach carries a structural limit.
| Method | What it does | Where it stops |
|---|---|---|
| De-hum | Places notches at a fundamental and its harmonics | Notch positions are locked to integer multiples of the fundamental. With a cap of 16 harmonics (7 in the plug-in build), a 200 Hz fundamental runs out at 3.4 kHz — it never reaches the hiss band above 6 kHz |
| Spectral de-noise | Estimates a noise profile and subtracts it | Low settings give musical noise, a chirpy watery ringing; high settings become broadband gating. The vendor documents this trade-off |
| Spectral repair | Makes a selection resemble its surroundings | Selection length is capped at 4-10 seconds depending on mode. A three-minute track means dozens of passes, and a floor that is always present has no clean surroundings to resemble |
| De-esser | Ducks the sibilance band dynamically | Sibilance is intermittent; the codec floor is continuous. A de-esser alone never crosses threshold on always-on content, so it needs a separate path beside it |
| Dynamic EQ, multiband | Ducks narrow bands conditionally | A good fit for stationary lines. But a human has to find and enter each target frequency, and repeat that per track and per generator |
| Regenerate or remaster | Rebuilds the song with a newer model | This is regeneration, not restoration, so the take you liked can change. And the output passes through the same class of neural codec, which leaves the same kind of residue again |
| Bury it in mastering | Raises loudness until it hides | Make-up gain lifts the noise floor too, and an air-band boost lands just above 9-12 kHz where the hiss is strongest, so it surfaces rather than hides |
The pattern is that the floor and the lines demand different tools. The floor needs statistical estimation and gating; the lines need stationary notches at located frequencies. Force one tool to do both and one of them goes too far. Chain two tools in series and their attenuations multiply, which is what collapses the top end.
Doing this by hand is also slow. Find the targets, enter the bands, repeat per stem, and A/B the result against the original — for every track.
3. How Honenuki removes it
Honenuki by Volta is an artifact remover built for AI music. It handles the two mechanisms on separate paths, and puts rules on how those paths combine.
| Path | Target | How | Cap |
|---|---|---|---|
| Hiss gate | High-frequency noise floor | Estimates a noise profile from a low time-percentile and gates it through an SNR ramp; time smoothing suppresses musical noise | 6-14 kHz, up to -40 dB |
| Standing-line notch | Non-harmonic stationary peaks | Confirms only lines that hold still over time and do not belong to a harmonic series, then notches them continuously | up to -20 dB |
| De-esser | Sibilance | Detects on the ratio of high-band to mid-band energy rather than an absolute threshold, so it does not drift with track level | separate path |
Design rules
- Masks combine with max, never by multiplication. Applying strength, de-esser and hiss in sequence stacks three attenuations on the same band and the top end collapses. Only the largest single attenuation is applied.
- The residual is computed before the restoration stage. Subtracting after saturation and air have been added contaminates what you hear as removed. You can only judge the work if the residual is exactly what was taken out.
- Listening and downloading share one path. What plays in the browser is the server render, and there is no preview-only processing.
- The same input with the same settings produces bit-identical output. This is subtraction, not regeneration, so the take you liked does not change.
Detection is display-only. Honenuki shows how much artifact it finds per stem, but that verdict never gates the processing — a false positive on human-made material does not block the work.
The strength knob is pinned to a displayed default of 40. Start there, listen to the residual, and move it up or down.
Questions
Isn't Suno's own Remaster enough?
It does a different job. Remaster re-generates the song with a newer model: quality can improve, but the take can change, and the output still comes out of a neural codec, so the high-frequency residue comes back with it. If you already like the take, fix that file rather than regenerating it.
Does splitting into stems remove it?
No. The artifacts are printed into the waveform at generation time, and separation happens afterwards. A separator divides components; it cannot undo residue that is already in them.
Won't cutting the top end make my mix dull?
It is not a broad high-frequency cut. Only narrow bands around detected peaks and the noise floor itself are reduced, and a restoration stage brings texture and brightness back. Drum stems use a lower attenuation cap and back off around transients, so attacks and cymbals survive.
Couldn't a better kernel choice prevent this at the source?
The paper thinks that is unlikely. Very particular kernel choices can mitigate the peaks, but the limited number of kernel parameters makes it unlikely to remove the artifacts entirely. Waiting for the generator side to solve it is a worse bet than post-processing.
Before or after mastering?
Before. Codec residue and high-frequency hiss get louder along with everything else when you raise the level, so clean first and master afterwards. Honenuki is not a mastering chain and applies no limiting or loudness normalisation.
Sign up and it is free during early access.
Sources and caveats
- Mechanism follows Afchar, Meseguer-Brocal, Akesbi and Hennequin (2025), A Fourier Explanation of AI-music Artifacts, arXiv:2506.19108, and Zeghidour et al. (2021), SoundStream: An End-to-End Neural Audio Codec, arXiv:2107.03312.
- The high-frequency over-smoothing description follows Kumar et al. (2023), High-Fidelity Audio Compression with Improved RVQGAN, arXiv:2306.06546.
- Those papers state explicitly that they cannot verify the internals of closed commercial models. They are cited here as evidence for the mechanism only; figures such as a specific generator's peak spacing do not come from them.
- The hiss band (6-14 kHz), the attenuation caps (-40 dB and -20 dB), and the observation that the noise floor concentrates around 8-12 kHz are Honenuki's own engine measurements, not values taken from a paper.
- The per-tool limits above follow the specifications published in each vendor's own documentation. This is not a ranking of products — it is a note on the structural constraints each design carries against this particular problem.