A Benchmark for Multi-Speaker Anonymization

Abstract

Privacy-preserving voice protection methods have mainly focused on single-speaker scenarios, which limits practicality in real multi-speaker conversations. We present an initial benchmark for multi-speaker anonymization, including task definition, evaluation protocol, baseline systems, and analysis of privacy leakage in overlapping speech. The benchmark systems combine speaker diarization with disentanglement-based anonymization and are further improved by conversation-level speaker-vector anonymization strategies that preserve speaker relations or increase anonymized speaker separability. Experiments on both simulated non-overlap and real-world datasets validate the effectiveness of the proposed system and anonymizers. We also provide analysis and potential solutions for overlapping speech leakage, along with code and evaluation data resources.