What AI music artifacts are, and how Honenuki measures and removes them
AI music artifacts are the residue a generator leaves behind in its output: faint but regular spectral peaks spaced across the frequency axis, plus a flat hiss floor in the high end. EQ and compression can mask them but cannot remove them, because the peaks sit inside the same bands as the music. Removing them takes a detector that locates the periodic grid and attenuates only the peaks that are actually there, narrow band by narrow band. Honenuki does that, then restores texture and top end with saturation, a de-esser and an air band.
Last reviewed 2026-09-02
Why Suno and other generators produce this sound
It is architectural, not a bug in any one model. Generative audio models build their output by upsampling — repeatedly expanding a compact internal representation back up to a waveform through transposed convolution (deconvolution) layers. Afchar and colleagues proved mathematically that these layers produce systematic frequency artifacts: small, distinctive spectral peaks, the audio relative of the checkerboard pattern familiar from image generators. Their key result is that the effect follows from the chosen architecture, not from the training data or the weights — which is why it survives every prompt, every genre and every re-roll. They confirmed it on open-source models and on the commercial generators Suno and Udio, and showed the peaks alone identify AI-generated music with over 99% accuracy in several settings. The same upsampling stack sits inside the neural audio codecs these systems decode through, which are built on residual vector quantization in the SoundStream line of work. That is the second half of the sound: a codec discards what it judges inaudible and reconstructs the rest, leaving a flat, static noise floor where detail used to be. So you get two things at once — a periodic peak structure that rings, and a hiss bed underneath it. Both survive stem export, and both are still there after ordinary mixing.
Mechanism and detection criterion: Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025), “A Fourier Explanation of AI-music Artifacts”, arXiv:2506.19108. Codec background: Zeghidour et al. (2021), “SoundStream: An End-to-End Neural Audio Codec”, arXiv:2107.03312.
How Honenuki works
Five rules govern the engine. They are worth stating plainly because they are what separate this from a de-noiser.
It finds the grid before it cuts
Detection looks for peaks that repeat at a regular spacing — around 200 Hz apart in the models we see most — and for a flat noise floor sitting just under the codec cutoff. Nothing is attenuated where no grid is found.
It cuts narrow, not wide
Attenuation is applied as narrow notches around the detected peaks only. There is no broadband high-shelf. Attenuation is capped at 8 dB inside the intelligibility band and 15 dB in the high band, so a mistake stays small.
The masks combine by maximum, never in sequence
Strength, de-ess and hiss each produce an attenuation mask. They are combined by taking the maximum at every point, not by multiplying one after another. Multiplying would compound into holes; taking the maximum means the deepest single reason wins and nothing stacks.
The residual is computed before anything is added
The REMOVED monitor is a true difference signal, captured before saturation and air are added back. It is what left, not what changed.
Drums are treated differently
Drum stems use a lower high-frequency attenuation cap and reduce attenuation around transients, so attacks and cymbals survive a setting that would otherwise dull them.
What each knob does
Six controls. Three take away, two put back, one is for your ears only.
Take away
Strength
How deep the notches go at the detected peaks. The display default is 40. Below a displayed value of about 2 the notch depth fades to zero, so a knob at the bottom is genuinely no processing rather than a little processing. Raising it deepens the cuts but never widens them.
Default 40 · caps 8 dB (intelligibility band) / 15 dB (high band)De-ess
A sibilance reducer for harsh S and T sounds. It is independent of artifact removal and works on any stem, but it defaults to 0.3 only on stems detected as vocal, and 0 elsewhere. It is not applied to the master, because de-essing a full mix damages everything that shares those frequencies.
Default 0.3 on vocal stems, 0 elsewhereHiss
The high-frequency hiss suppressor, working between 6 and 14 kHz with a cosine taper at the top so nothing changes abruptly at the boundary. It combines a dynamic noise gate that opens and closes with the signal-to-noise ratio, and a standing comb notch for the fixed peaks. The gate reaches 40 dB of attenuation at the maximum setting; the comb notch is capped at 20 dB.
6–14 kHz · gate up to −40 dB · comb notch up to −20 dB · default knob 0.1 = −4 dBPut back
Saturation
Three-band harmonic generation that puts density back after removal. The value is a target distortion figure (THD), not a gain, and the preset — WARM, PRESENT or AGGR — decides which bands are driven and the balance of even to odd harmonics. It defaults to 0, so nothing is added unless you ask.
Default 0 · value is target THD % · presets WARM / PRESENT / AGGRAir
A restoration shelf for the top end that removal takes with it. The readout is the net change against the original, so 0 dB means you are back where you started, and above that you are adding brightness the source did not have. It does not resynthesize frequencies the codec discarded — it lifts what remains.
−2 to +20 dB · default +1.0 dBListening only
Gain
Monitoring level only. It changes what you hear while comparing and is deliberately excluded from the render, so nothing you do here can end up in the downloaded file.
−24 to +12 dB · never applied to the downloadWhat Honenuki does not do
- It is not a stem separator. It cleans the file you upload, whether that is a full mix or a stem you already have. It will not split a song into vocals, drums and bass.
- It is not a watermark remover. It does not touch inaudible watermarks such as SynthID or C2PA credentials, and it makes no claim about distributor or streaming-platform screening.
- It is not a mastering chain. There is no limiting and no loudness normalisation. Clean first, master afterwards.
- It is not a bandwidth extender. It does not invent frequencies the codec threw away.
- It is not forensic proof. The AI verdict reports codec fingerprints for your information. It never blocks processing, and a stem judged human-made is treated exactly as you ask.
Figures
| Item | Value |
|---|---|
| Detected grid spacing | approximately 200 Hz |
| Hiss suppression band | 6–14 kHz, cosine taper at the top |
| Hiss gate depth | up to −40 dB at knob 1.0 (default 0.1 = −4 dB) |
| Standing comb notch cap | −20 dB |
| Notch caps | 8 dB intelligibility band, 15 dB high band |
| Strength default (displayed) | 40 |
| Air range / default | −2 to +20 dB / +1.0 dB |
| Mask combination | maximum, never sequential multiplication |
| Residual | computed before saturation and air are added |
| Input formats | WAV, MP3, FLAC, AIFF |
| Output | WAV (also MP3), delivered as one ZIP |
| Determinism | same input and settings give a bit-identical file |
| Listening | the browser plays the same server render you download |
Questions
Can I just use an EQ instead?
Not for this. The peaks sit inside the same bands as the music, so a static EQ cut wide enough to catch them takes the music with it. What works is a narrow notch placed where a peak actually is, which means finding the grid first.
Does a higher strength always sound better?
No. Strength deepens the notches; past the point where the peaks are gone it starts eating the material around them. The useful test is the REMOVED monitor — if you hear melody or words in it, the setting is too high.
Why is de-ess off on non-vocal stems?
Because sibilance is a vocal problem, and de-essing a drum or synth stem attenuates content that shares those frequencies for no benefit. The knob is still available on any stem if you want it.
Why is there no de-ess knob on the master?
De-essing a full mix acts on every source that shares those frequencies, not just the voice. De-ess belongs on the vocal stem, which is why it appears in per-stem editing rather than on the master.