Brief

Researchers have developed a new training framework called RIVET to improve the robustness of voice attribute editing models. This framework incorporates an idempotency objective, which ensures that repeated application of an editing function yields the same result, thereby reducing sensitivity to noisy or inconsistent attribute annotations in large-scale speech datasets. Evaluations on controlled label noise and the GLOBE dataset demonstrate that RIVET enhances editing success and better preserves speaker identity compared to standard training methods. AI