Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Third VoicePrivacy Challenge: Preserving Emotional Expressiveness and Linguistic Content in Voice Anonymization

Natalia Tomashenko, Xiaoxiao Miao, Pierre Champion, Sarina Meyer, Michele Panariello, Xin Wang, Nicholas Evans, Emmanuel Vincent, Junichi Yamagishi, Massimiliano Todisco

TL;DR

The paper reports on the VPC 2024, a competition advancing voice anonymization to conceal speaker identity while preserving linguistic content and emotional expressiveness. It documents a diverse suite of methods—from neural codecs and attribute disentanglement to cascaded ASR+TTS and hybrid approaches—evaluated under a semi-informed attacker model using EER for privacy and WER/UAR for utility. The results show neural codecs and VC-based systems delivering strong privacy-utility balance, while cascaded systems achieve high privacy at the expense of emotion preservation; hybrid methods offer flexible trade-offs. The work underscores challenges in reliable privacy assessment, the value of multi-attribute disentanglement, and the need for richer datasets and robust evaluation as the field moves toward future challenges and real-world deployment.

Abstract

We present results and analyses from the third VoicePrivacy Challenge held in 2024, which focuses on advancing voice anonymization technologies. The task was to develop a voice anonymization system for speech data that conceals a speaker's voice identity while preserving linguistic content and emotional state. We provide a systematic overview of the challenge framework, including detailed descriptions of the anonymization task and datasets used for both system development and evaluation. We outline the attack model and objective evaluation metrics for assessing privacy protection (concealing speaker voice identity) and utility (content and emotional state preservation). We describe six baseline anonymization systems and summarize the innovative approaches developed by challenge participants. Finally, we provide key insights and observations to guide the design of future VoicePrivacy challenges and identify promising directions for voice anonymization research.

The Third VoicePrivacy Challenge: Preserving Emotional Expressiveness and Linguistic Content in Voice Anonymization

TL;DR

The paper reports on the VPC 2024, a competition advancing voice anonymization to conceal speaker identity while preserving linguistic content and emotional expressiveness. It documents a diverse suite of methods—from neural codecs and attribute disentanglement to cascaded ASR+TTS and hybrid approaches—evaluated under a semi-informed attacker model using EER for privacy and WER/UAR for utility. The results show neural codecs and VC-based systems delivering strong privacy-utility balance, while cascaded systems achieve high privacy at the expense of emotion preservation; hybrid methods offer flexible trade-offs. The work underscores challenges in reliable privacy assessment, the value of multi-attribute disentanglement, and the need for richer datasets and robust evaluation as the field moves toward future challenges and real-world deployment.

Abstract

We present results and analyses from the third VoicePrivacy Challenge held in 2024, which focuses on advancing voice anonymization technologies. The task was to develop a voice anonymization system for speech data that conceals a speaker's voice identity while preserving linguistic content and emotional state. We provide a systematic overview of the challenge framework, including detailed descriptions of the anonymization task and datasets used for both system development and evaluation. We outline the attack model and objective evaluation metrics for assessing privacy protection (concealing speaker voice identity) and utility (content and emotional state preservation). We describe six baseline anonymization systems and summarize the innovative approaches developed by challenge participants. Finally, we provide key insights and observations to guide the design of future VoicePrivacy challenges and identify promising directions for voice anonymization research.
Paper Structure (83 sections, 3 equations, 13 figures, 7 tables)

This paper contains 83 sections, 3 equations, 13 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Challenge design
  3. Voice anonymization task
  4. Attack model
  5. Data and pretrained models
  6. Training resources
  7. Development and evaluation data
  8. Privacy and utility evaluation
  9. Training data for evaluation models
  10. Privacy metric: the equal error rate (EER)
  11. Utility metrics
  12. Word error rate (WER)
  13. Unweighted average recall (UAR)
  14. Privacy-utility tradeoff
  15. Baseline voice anonymization systems
  16. ...and 68 more sections

Figures (13)

  • Figure 1: Privacy preservation scenario as a game between users and attackers in the case where speaker identity is considered as personal information to be protected using anonymization, while linguistic and emotional content should be preserved for utility downstream tasks. Privacy evaluation of anonymized data is performed using a semi-informed attack model $ASV_\text{eval}^{\text{anon}}$ and equal error rate (EER) metric. Utility evaluation is performed using (1) automatic speech recognition (ASR) model $ASR_\text{eval}$ and word error rate (WER) metric and (2) speech emotion recognition (SER) model $SER_\text{eval}$ and unweighted average recall (UAR) metric.
  • Figure 2: Example system rankings according to the privacy (EER) and utility (WER and UAR) results for 4 minimum target EERs. Different colors correspond to 6 different teams. Numbers within each circle show system ranks for a given category. Grey circles correspond to the baseline systems, and the black circle to the original (unprocessed) data.
  • Figure 3: Baseline anonymization system B1.
  • Figure 4: Baseline anonymization system B3.
  • Figure 5: Baseline anonymization system B4.
  • ...and 8 more figures