Reproduce “Efficient Speech Watermarking for Speech Synthesis via Progressive Knowledge Distillation” 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
Yang Cui, Peter Pan, Lei He, Sheng Zhao
Microsoft, Beijing, China, researchers developed PKDMark, a post-hoc speech watermarking method that uses progressive knowledge distillation to achieve high robustness against diverse audio attacks and real-time computational efficiency. This approach reduces processing costs by 93.6% while maintaining speech imperceptibility (PESQ 4.30) and robust watermark detection (average BER 0.51%).
Key insights
- A complex domain message feature map and a multi-scale STFT loss within a multi-objective joint optimization are crucial for the teacher model's superior robustness and perceptual quality.
- Progressive Knowledge Distillation effectively enables a compact student model to inherit strong robustness from a complex teacher, outperforming direct knowledge distillation.
- The framework successfully reconciles the trade-off between robustness and computational efficiency, making deep learning-based speech watermarking viable for real-time applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.