This paper presents a new benchmark to quantitatively evaluate the residual identity leakage problem of speaker anonymization techniques that conceal speakers' identities while maintaining their identifiable information. We measure residual identity leakage using three error rates: the Equal Error Rate, the Cumulative Match Characteristic hit rate, and embedding space similarity using Canonical Correlation Analysis and Procrustes Analysis. Our evaluation results show that all state-of-the-art speaker anonymization systems leak identity information, with even the best performing system only slightly better than random guessing, and the worst performing system achieving a 45% accuracy rate among the top 50 speakers based on CMC. This highlights the persistent privacy risks of current speaker anonymization techniques.
