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. An auxiliary speaker classification branch with a gradient reversal layer is used during training to discard speaker-specific information. Results show that this straightforward embedding-based approach achieves a very low word error rate (WER 2.53) with an anonymization performance (EER 13.39) ranking within the 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.
Blogger's Review: This research successfully implements an efficient voice anonymization technique through innovative acoustic processing methods, showcasing the potential of content embeddings in speech processing. The preservation of emotional information particularly enhances its applicability, suggesting that future developments could significantly improve voice privacy protection effectiveness.