Web Analytics
This site demonstrates one possible use of this domain. For acquisition, partnership, or investment inquiries, please use our contact form.

Security · MIT CSAIL · 2026

Detection Limits of State-of-the-Art Neural Voice Clones

Park, Iyer, BauerOpen-source detectors trained in 2024 fall to under 60% accuracy on 2026-era commercial voice clones, suggesting a structural disadvantage for post-hoc detection.

Key findings

  • Cross-model detection generalization degrades by an average of 22 points within 18 months of a new generator release.
  • Provenance watermarks survive 88% of common transformations but only 14% of adversarial re-encodes.

What it means in practice

Reliable provenance has to be established at capture time, not inferred from the artifact later. Detection alone is a losing strategy.