Voice & Expression Identity Lock

Music that can’t be used to train a model.

VEIL protects recorded music from being used as training data by generative AI. An artist runs a track through it before release; the track sounds the same, but a model that trains on it doesn’t learn how to sound like them.

The problem: Text-to-music models learn from enormous collections of recorded music, mostly scraped without anyone asking. Once a model has learned an artist, anyone can generate new tracks in that artist's style, their voice included, for the price of a prompt.
What VEIL does: Before a track is released, VEIL adds a perturbation to it that listeners should not be able to hear. A generative model trained on the protected audio fails to learn the artist's musical identity: their sound, their voice, their way of playing. Its imitations drift away from the artist and toward junk.
Why it works: The perturbation is computed, not random. Training walks straight through noise the ear can't hear. VEIL solves for it against the same audio encoders real generators are built on, shapes it per stem so loud drums carry more of it than an exposed vocal, and hardens it against compression, resampling and source separation.