Reproduce “DiffAnon: Diffusion-based Prosody Control for Voice Anonymization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ismail Rasim Ulgen, Zexin Cai, Nicholas Andrews, Philipp Koehn, Berrak Sisman

A diffusion-based voice anonymization framework, DiffAnon, provides explicit and continuous inference-time control over source prosody preservation using classifier-free guidance. This enables dynamic adjustment of the utility–privacy trade-off within a single model, demonstrating a controlled decrease in F0 correlation from 75.58% to 62.45% while improving privacy EER from 33.09% to 42.43%.

Key insights

  • Prosody serves as a principal and controllable determinant of the utility–privacy trade-off in voice anonymization.
  • Diffusion models, when conditioned on multi-level speech codec embeddings like those from RVQ, can enable fine-grained, controllable generation of acoustic attributes for anonymization.
  • Classifier-free guidance can be effectively adapted to allow continuous, inference-time adjustment of prosody preservation and anonymization strength within a single pre-trained model.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.