Design Choices for X-Vector Based Speaker Anonymization
Brij Mohan Lal Srivastava, Natalia Tomashenko, Xin Wang, Emmanuel Vincent, Junichi Yamagishi, Mohamed Maouche, Aurélien Bellet, Marc Tommasi
Abstract
The recently proposed x-vector based anonymization scheme converts any input\nvoice into that of a random pseudo-speaker. In this paper, we present a\nflexible pseudo-speaker selection technique as a baseline for the first\nVoicePrivacy Challenge. We explore several design choices for the distance\nmetric between speakers, the region of x-vector space where the pseudo-speaker\nis picked, and gender selection. To assess the strength of anonymization\nachieved, we consider attackers using an x-vector based speaker verification\nsystem who may use original or anonymized speech for enrollment, depending on\ntheir knowledge of the anonymization scheme. The Equal Error Rate (EER)\nachieved by the attackers and the decoding Word Error Rate (WER) over\nanonymized data are reported as the measures of privacy and utility.\nExperiments are performed using datasets derived from LibriSpeech to find the\noptimal combination of design choices in terms of privacy and utility.\n