Speaker Anonymization

Speaker anonymization studies how to make speech harder to link to a person while retaining what listeners and downstream systems need, including words, emotion, and clinical cues. Common approaches use voice conversion, pseudo-speaker embeddings, self-supervised or codec representations, diffusion models, and text-to-speech; streaming systems add constraints on latency and future context. Evaluation compares privacy against intelligibility, speech quality, and task performance, using speaker-verification attacks alongside measures of linguistic and attribute leakage. A central tension is that privacy depends on the attacker, anonymizer, speaker, and available speech, so strong performance on one metric or benchmark may not ensure robust protection or preserve every useful attribute.