AI Audio Detector — Tell if a track is AI-generated

Honestly: heuristic-only detection (Quick mode) struggles to reliably tell modern frontier AI generators from clean studio recordings. That is why we return "Inconclusive" when the heuristics cannot make a confident call — instead of guessing 40 %. Deep mode improves accuracy by analyzing vocal and instrumental stems separately, and a Phase C ML classifier (in development) will raise this further. We publish our limits openly.

Deep mode separates your track into vocal and instrumental stems (using the same UVR model as our /vocal-remover) and analyses each one independently. This catches AI-cloned vocals on real instrumentals (and vice versa) — a signature that single-mix detectors miss completely. Stem disagreement itself is a positive signal we surface as a "Mixed" verdict.

Because lying with false precision is worse than admitting uncertainty. When the 12 heuristics produce widely-spread votes OR the average sits near the boundary, we say so explicitly. This was a fix for user feedback: previous versions returned ~40 % to everything, which was misleading. "Inconclusive" tells you to trust your ear or run Deep mode.

Heuristic patterns common to all modern AI music and voice generators — both song-generation and voice-cloning systems. Detection is based on acoustic properties (regularity, spectral signatures, training-corpus cutoffs) plus a separate watermark / metadata layer that catches generators which voluntarily tag their output. The architecture stays valid as new generators emerge.

Processing runs in your browser via WebAssembly — analysis itself doesn't require an upload. Standard server-side telemetry (page views, error events) applies as on any web page; see Privacy Policy for details.

Yes — Deep mode's vocal stem analysis is the right tool. Cloned voices typically show AI-tells in the vocal stem while the surrounding music can be either AI or human. The "Mixed" verdict surfaces this directly.