Velma Triage

Every window is checked independently for AI vocals and AI instrumentals — a track can flag on both at once. A dash means that detector had nothing to score in the window; a low score means it checked and the content looks real. Heavily processed or high-production songs may still false-positive.

Whole-clip scores over a fixed set of 42 sound-event labels — crowd reactions, animals, instruments, household sounds. Each score is an independent probability, so several events can be detected at once. Plain speech has no label of its own: a talk-only clip scores low everywhere. Events at 50% or higher count as detected.

The model turns each clip into a compact voice fingerprint and scores how similar the two fingerprints are, from −1 to 1. It compares how a voice sounds, not what is said — the same speaker saying different things still matches. The verdict is set by this showcase, not the API; treat scores near a boundary as inconclusive. Recording conditions matter: two speakers on the same call score closer than two speakers from unrelated recordings.