The article discusses speaker anonymization, a method to protect a speaker’s identity without degrading speech quality and intelligibility. The authors propose SALT, a Speaker Anonymization system based on Latent space Transformation. This system extracts latent features via a self-supervised feature extractor and then interpolates the latent vectors to achieve speaker anonymization. This method aims to address issues with existing systems that can reduce speaker distinctiveness, speech quality, and intelligibility for some speakers. The authors suggest their system offers state-of-the-art distinctiveness while preserving speech quality and intelligibility.
Publication date: 10 Oct 2023
Project Page: https://github.com/BakerBunker/SALT
Paper: https://arxiv.org/pdf/2310.05051
