RVCBench from UBC stress-tests 18 voice-cloning models under real-world noise and attacks
机器之心 · wechat · 2026-10-21
Researchers from UBC and Vector Institute built RVCBench, a NeurIPS 2026 benchmark decomposing voice-cloning robustness into input, generation, output and audio-perturbation dimensions across 10 task types and 14,370 samples, evaluating 18 open-source models.
Key findings: clean-testset scores don't predict robustness (StyleTTS2 has 0.01 WER but weak speaker similarity); bigger isn't more robust—639M CosyVoice beats 5.97B Higgs in multilingual settings; reference audio of 8–12s is the sweet spot; background noise degrades WER by 265.6% relatively; anti-cloning protections like SafeSpeech and SPEC cut similarity from 0.492 to 0.22–0.24 and survive naive denoising. The benchmark ships as a pip-installable toolkit (rvcbench) with a 22-model leaderboard.
More from Research
- OPAL: an LLM agent rewrites the optimizer's source code during training — RobertTLange · 2026-10-23
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23