RIVET: Robust Idempotent Voice Attribute Editing

The article introduces RIVET, a training framework for voice attribute editing models that incorporates an idempotency objective to enhance robustness against noisy label annotations. By enforcing idempotency, the framework reduces sensitivity to mislabeled data, leading to improved editing success and better preservation of speaker identity when evaluated on the GLOBE dataset. This approach is significant for practitioners as it addresses common challenges in training generative models on large-scale datasets with inconsistent attribute labels, ultimately enhancing the reliability of voice editing applications.