LiveAnon: Improving Privacy and Utility in Streaming Speaker Anonymization with Codec Language Models
LiveAnon is a voice anonymization framework that uses a streaming Automatic Speech Recognition (ASR) encoder to extract content features from source speech. These features condition an autoregressive neural audio codec, which generates target speech conditioned on the source content and a reference prompt. This method preserves emotion information without needing specific emotion supervision, achieving strong privacy and utility balance.