Guide · Mastering
Typical artefacts in AI music – and how to tame them
Music generators deliver impressive results, but they have a sonic fingerprint of their own. Once you know it, you hear it immediately – and can soften it. Here are the five artefacts we ran into most often across more than 50 released tracks, with what helped each time.
1. Sibilance and “glitter” between 5 and 9 kHz
How to spot it: S and Z sounds jump out, hi-hats sound like breaking glass, long listening becomes tiring. More obvious on headphones than on speakers.
Cause: the models work internally with compressed representations of the spectrum; fine highs are often overdrawn during reconstruction.
What helps: a de-esser that only steps in when the high band gets loud – which is exactly what the “AI artefact cleanup” in the mastering tool does. A static EQ cut, by contrast, also removes air and sheen where you want them. In the prompt, words like “warm”, “analog”, “tape”, “soft highs” help; “crisp”, “bright” and “modern” make the problem worse.
2. Rumble below 30 Hz
How to spot it: you can’t – on most speakers it is inaudible. But it eats headroom: the true peak rises, the limiter fights something nobody hears, and the whole song ends up quieter.
What helps: a gentle low cut at 30 Hz (second order, no steep filter). In the tool it is part of the artefact cleanup. For bass-heavy genres use the “Normal” setting, not “Strong”.
3. The phase cloud
How to spot it: the song sounds huge in stereo, but as soon as you switch to mono (phone speaker, Bluetooth box) it collapses: vocals quieter, guitars vanish, everything sounds hollow.
Cause: generators often create stereo width through phase shifting rather than real space.
What helps: listen to the song in mono once before releasing. If it collapses, the only mastering fix is slightly reducing the width above about 200 Hz – a tool for this is planned for a later version. More sustainable is a new run with “mono-compatible”, “centered vocals” or “intimate” in the prompt.
4. Mushy consonants and “rubber syllables”
How to spot it: the lyrics are unintelligible in places, syllables merge, and at line endings sounds appear that do not exist in the language. More frequent in German lyrics than in English ones.
Cause: the model knows the language less well, or the line has too many syllables for the melody.
What helps: no mastering in the world. Instead: shorter lines, clear vowels at line endings, punctuation as breathing pauses, capital-letter emphasis only where it belongs. And several runs – the hit rate rises with every attempt. Our experience: three to six generations per German song until one version comes out without rubber syllables.
5. The “pumping” effect and too much compression
How to spot it: after every kick the whole mix breathes in, pads and vocals briefly duck and come back. Especially in electro and dance.
Cause: the models learned from fully mastered songs and bring compression with them. Another strong compressor in mastering doubles it.
What helps: mastering that mainly places the song rather than densifying it again. That is why the tool’s profiles use low ratios (1.3:1 to 2.2:1) and long release times. If it still pumps, the original is already too dense – then the “Neutral” profile is the better choice.
The order for checking
- Listen to the song on three systems: headphones, laptop speakers, phone. Artefacts 1, 3 and 4 each show up on a different one.
- Use the mono switch if your player has one.
- Improve the prompt and generation first, then master. Mastering is the last step, not the repair.
An AI song where nobody hears the artefacts any more is no accident – it is the result of ten discarded runs and a restrained master.
All guides are deliberately kept general: which models and tools run on your machine depends on your hardware. Questions or corrections? Write to us.