Abstract

We present a novel approach to de-anonymize speech that has been transformed by a voice privacy system. Inspired by the complex and multi-factorial nature of speaker identification, we extract three different identity-related features, namely X-Vectors, pitch-based representations, and prosody embeddings. These features are then fused together and used to perform speaker verification on the anonymized data. By integrating multiple parallel streams of identity information, we increase the robustness of the system to different voice conversion methods and also allow for easy fine-tuning to exploit the unique weaknesses of a specific anonymization method.

Date

November 30, 2025

Authors

Thomas Thebaud, Nicholas Mehlman, Yaohan Guan, Laureano Moro-Velazquez, Jesus Villalba Lopez, Shrikanth Narayanan, Najim Dehak

Conference

Proc. SPSC 2025

Pages

61-67