Generative music is now good enough that catalogues, labels and platforms need to know what they are looking at. Audio AI Remover answers one question: was this made by a model or by people?
The current V3 model is a gradient-boosted classifier over a purpose-built feature set, trained on a human/AI corpus. It scores 0.94 accuracy in validation and holds 0.89 on a blind evaluation set — the number that matters, because it is the one measured on material the model had never encountered in any form.
The model is trained and evaluated; it has not yet been promoted to production, which is deliberate — a detector that is wrong in public is worse than no detector.
At a glance
- Model
- V3 · gradient-boosted classifier
- Accuracy
- 0.94 validation
- Blind set
- 0.89
- Status
- Trained, not yet in production
What it does
Binary verdict
Human or AI, with a confidence score rather than a bare label.
Measured honestly
0.94 validation accuracy, 0.89 on a blind set — both reported, not just the flattering one.
Feature-based, not a black box
A purpose-built feature set over a gradient-boosted model, so decisions can be inspected.