You Are What You Say: Exploiting Linguistic Content for VoicePrivacy Attacks

TL;DR

This paper demonstrates that linguistic content can significantly compromise speaker anonymization systems, revealing privacy vulnerabilities through adapted language models and suggesting dataset improvements.

Contribution

It introduces a BERT-based ASV attack that exposes linguistic content vulnerabilities in speaker anonymization, highlighting the need for better dataset design and evaluation metrics.

Findings

Achieved a mean EER of 35% using linguistic content-based attack.

Certain speakers had EERs as low as 2%, indicating high privacy risk.

Revealed that dataset curation influences attack success and evaluation fairness.

Abstract

Speaker anonymization systems hide the identity of speakers while preserving other information such as linguistic content and emotions. To evaluate their privacy benefits, attacks in the form of automatic speaker verification (ASV) systems are employed. In this study, we assess the impact of intra-speaker linguistic content similarity in the attacker training and evaluation datasets, by adapting BERT, a language model, as an ASV system. On the VoicePrivacy Attacker Challenge datasets, our method achieves a mean equal error rate (EER) of 35%, with certain speakers attaining EERs as low as 2%, based solely on the textual content of their utterances. Our explainability study reveals that the system decisions are linked to semantically similar keywords within utterances, stemming from how LibriSpeech is curated. Our study suggests reworking the VoicePrivacy datasets to ensure a fair and unbiased evaluation and challenge the reliance on global EER for privacy evaluations.