Publikation
A Comprehensive Evaluation Framework for Speaker Anonymization Systems
Carlos Franzreb; Tim Polzehl; Sebastian Möller
In: Proc. 3rd Symposium on Security and Privacy in Speech Communication. Symposium on Security and Privacy in Speech Communication (SPSC-2023), located at Interspeech 2023, August 21-24, Dublin, Ireland, ISCA, 2023.
Zusammenfassung
Speaker anonymization consists of concealing the source speaker’s identity, while keeping the linguistic and paralinguistic content intact. It is usually evaluated as a trade-off between privacy and utility. The current standard for privacy evaluation is the automatic speaker verification system (ASV) from the Voice Privacy Challenge (VPC); it involves computing speaker embeddings and comparing them with a trained PLDA algorithm. We implement this ASV system and extend the utility evaluation of the VPC, which previously consisted of automatic speech recognition performance, with emotion preservation, naturalness and performance metrics. Our framework is fast and easily customizable, facilitating the development and evaluation of new anonymization pipelines. We showcase this framework with the StarGANv2-VC, one of the most powerful voice conversion systems available.