arXiv cs.AI· Adrien Schneider (M-PSI), Kacper Zabkowski (M-PSI), Anderson Augusma (M-PSI), Fr{\'e}d{\'e}rique Letu{\'e} (SVH), Maria Camila Pinzon (M-PSI), Dominique Vaufreydaz (M-PSI)·· 10 小时前AI 评分29
基于内容嵌入匹配的语音匿名化模型：跳过信号重建，WER 2.53、EER 13.39
Listen to the Features: Voice Anonymization Driven by Content Embedding Matching over Signal Reconstruction
AI 导读
一项研究提出以内容保留为目标的语音匿名化模型，用冻结的预训练 wav2vec2 编码器提取内容嵌入，经向量量化和 HiFi-GAN 声码器解码为匿名语音，训练中不设波形重建损失、不做说话人嵌入映射。
正文 · 原文
Abstract:The paper presents a voice anonymization model focusing on preserving content rather than producing realistic speech. It relies on content embeddings extracted from a frozen pretrained wav2vec2 encoder. These embeddings are decoded into an anonymized signal using vector quantization and a HiFi-GAN vocoder, both trained on LibriTTS without any waveform reconstruction loss or speaker embedding mapping. The training objective enforces that embeddings of the anonymized signal match those of the original one. While training, an auxiliary speaker classification branch with a gradient reversal layer is used to discard speakerspecific information. Results show that this straightforward embedding-based approach achieves very low WER (2.53) with an anonymization performance (EER 13.39) ranking within first level for VPC. Notably, emotions are partially preserved (UAR 43.91), even without a supporting training objective, while the anonymized voice is audible without reconstruction loss.
| Subjects: | Signal Processing (eess.SP); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.09767 [eess.SP] |
| (or arXiv:2607.09767v2 [eess.SP] for this version) | |
| https://doi.org/10.48550/arXiv.2607.09767 arXiv-issued DOI via DataCite |
|
| Journal reference: | 6th Symposium on Security and Privacy in Speech Communication joined with 4th Voice Privacy Challenge at Interspeech 2026, Sep 2026, Sidney, NSW, Australia |
Submission history
From: Dominique Vaufreydaz [view email] [via CCSD proxy]
[v1]
Tue, 7 Jul 2026 09:11:20 UTC (158 KB)
[v2]
Tue, 10 Nov 2026 15:25:41 UTC (161 KB)
来源：arXiv cs.AI · arxiv.org