Reproduce “VocBulwark: Towards Practical Generative Speech Watermarking via Additional-Parameter Injection” 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
Weizhi Liu, Yue Li, Zhaoxia Yin
VocBulwark introduces a generative speech watermarking framework that injects watermarks via additional, lightweight parameters to maintain high perceptual fidelity while achieving robust traceability for AI-generated speech. This method demonstrated resilience against diverse attacks, including neural codecs and variable-length manipulations, maintaining over 99% accuracy at 2000 bps on in-distribution data and strong generalization to out-of-distribution datasets.
Key insights
- Injecting watermarks through auxiliary, lightweight parameters rather than altering the main generative model's frozen weights effectively preserves speech perceptual quality while enabling deep embedding.
- Frame-level temporal broadcasting and acoustic feature alignment in the embedding module significantly enhance watermark resilience against variable-length manipulations and preserve intrinsic acoustic attributes.
- A curriculum-guided training strategy is crucial for harmonizing the conflicting objectives of maintaining high speech fidelity and achieving robust watermark extraction, ensuring stable convergence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.