Diagnose before adding another effect
An AI music artifact is an unwanted audible result, not a single kind of noise with one universal fix. Describe the sound, mark its timestamp, and identify the first version where you can hear it. Compare the original download with the separated stem, processed mix, and export at similar playback levels.
If the issue begins after a processing stage, return to that stage. If it exists in the source, decide whether a local repair, another take, or a different arrangement would best preserve the performance. Keep the original file untouched throughout.
Follow the closest symptom
- Metallic vocals: ringing vowels, brittle consonants, or a hollow edge that needs careful tonal treatment.
- Chirps and warble: brief glitches or sustained watery notes that need different diagnosis and editing decisions.
- Stem cleanup: leakage and damaged detail that should be judged with the accompaniment restored.
These guides offer troubleshooting methods, not measured failure rates for a model. A problem in one song does not establish that every song made by the same generator has it.
Audacity’s spectrogram documentation explains how the view represents sound across time and frequency. Use it to locate an event you hear. Do not erase every bright shape: the wanted performance creates shapes too.
Keep a fair stopping rule
Compare a repaired problem passage, an unchanged clean passage, and a dense section. Match loudness before choosing. Reject a repair that makes the defect quieter but removes lyric clarity or musical energy.
Once the track works during ordinary listening, continue to mixing and mastering. Do not keep processing merely to make a display look cleaner. Audible defects and embedded provenance are separate questions; the watermarks and detection hub explains what each kind of evidence can establish.