Researchers Built a Tool That Makes Your Vocals Uncloneable

Researchers at Binghamton University just quietly changed the terms of the voice cloning fight, and most of the industry hasn't caught up yet.

The tool, recently presented at the 39th NeurIPS Workshop, is called My Music My Choice — MMMC — built in collaboration with the startup Cauth AI.

It works by embedding imperceptible alterations into a song's audio waveform: changes the human ear never registers, but that are enough to corrupt a generative AI model's ability to clone what it hears.

When an AI attempts to replicate a protected vocal, the output comes back distorted and unusable. The clone fails at the source. Testing on 150 tracks across genres confirmed the approach holds, and the research landed at NeurIPS 2025, one of machine learning's most competitive venues. The community is treating this as real infrastructure.

What makes MMMC structurally interesting isn't just what it does. It's where the intervention happens. Most protection efforts in this space have focused on detection after the fact.

Platforms scan for AI-generated audio, distributors flag suspicious uploads, labels play whack-a-mole on streaming services. Companies like IRCAM Amplify, Deezer, Pex, and CoverNet have built a growing B2B detection economy targeting exactly this problem. That infrastructure matters. But it places the burden of defense downstream, after the damage is already done.

MMMC inverts that logic. It turns the original recording into a trap. An artist doesn't need to wait for a clone to surface and then issue a takedown because the protection is already baked into the file before it ever leaves their hands. This is closer to how physical locks work than how surveillance systems work. Preemptive, not reactive.

The real tension here is the arms race. Researchers have already demonstrated that music deepfake detection faces serious challenges around calibration, robustness to audio manipulation, and generalization across model types, and the same pressure will eventually apply to MMMC. As voice cloning models grow more sophisticated, the question is whether waveform-level interventions can evolve fast enough to stay ahead of what they're designed to stop. A determined bad actor with access to the underlying model architecture could, in theory, train around the perturbation.

But that's a problem for version two. What matters right now is the framing shift. The research community is beginning to treat artist voice protection as an engineering problem with solvable components, not just a legal or policy problem that requires industry coordination. That distinction is significant, because it means the tools might actually get built.

For independent artists especially, a file-level defense that travels with the recording itself is qualitatively different from anything currently on the market. The voice is the asset. Protecting it at the waveform is, finally, protecting it at the source.

Source: Binghamton University / Cauth AI, presented at NeurIPS 2025 Workshop: AI for Music