[論文レビュー] The Third VoicePrivacy Challenge: Preserving Emotional Expressiveness and Linguistic Content in Voice Anonymization

Tomashenko, Natalia, Xiaoxiao Miao+8 more|arXiv (Cornell University)|Jan 17, 2026

Speech Recognition and Synthesis被引用数 0

ひとこと要約

The paper presents the 2024 VoicePrivacy Challenge, detailing task setup to anonymize speaker identity while preserving linguistic content and emotional state, along with datasets, attack model, evaluation metrics, baselines, and 36 participating systems.

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.

研究の動機と目的

• Motivate privacy-preserving speech processing under GDPR-like constraints by concealing speaker identity.
• Preserve linguistic content and emotional state to maintain downstream utility for ASR and SER tasks.
• Describe challenge framework, attack model, datasets, and evaluation metrics to benchmark anonymization methods.
• Present baseline systems and analyze participant approaches to guide future VoicePrivacy research.

提案手法

• Define an utterance-level anonymization task that replaces speaker identity with a pseudo-speaker while keeping content and emotion intact.
• Adopt a semi-informed attack model where an attacker uses anonymized enrollment to re-identify speakers via ASV.
• Use LibriSpeech and IEMOCAP data for development/evaluation; train ASV/ASR/SER models on standard corpora to assess privacy and utility.
• Evaluate privacy with EER improvements of ASV on anonymized data; evaluate utility with WER for ASR and UAR for SER.
• Provide six baseline anonymization systems (B1–B6) and summarize 36 submitted systems with diverse approaches.

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

実験結果

リサーチクエスチョン

• RQ1 Can speaker identity be effectively hidden in utterances while preserving linguistic content and emotional state?
• RQ2 How do different anonymization strategies balance privacy (higher EER) and utility (lower WER, higher UAR)?
• RQ3 What are the strengths and limitations of current baselines and participant approaches for VPC 2024?
• RQ4 How does stronger attacker modeling influence privacy evaluation and protocol design for future challenges?

主な発見

• The 2024 edition extends prior work by requiring preservation of emotional state in addition to linguistic content.
• Evaluation uses EER for privacy, and WER/UAR for utility, under a semi-informed attack model.
• Six baselines and 36 submitted systems demonstrate diverse approaches, including neural vocoders, GAN-based anonymization, neural codecs, and ASR/BN with VQ techniques.
• Results highlight trade-offs between privacy guarantees and downstream task performance, informing future challenge design and research directions.

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