Reproduce “AudioNoisePrints: Model-free audio watermarking using spatial correlation in flow matching TTS” on OpenResearch

About this paper

Timothy Tin-Long, Jian Zhu, Aidan Pine, Mengzhe Geng

AudioNoisePrints introduces a training-free watermarking pipeline for flow matching and diffusion Text-to-Speech models by leveraging the inherent spatial correlation between initial Gaussian noise and the generated audio's Mel-spectrogram. The method achieves robust detection against various audio augmentations, notably cropping and speed scaling, without requiring retraining the generative model or adding computational overhead during inference.

Key insights

• Strong spatial correlations exist between the initial Gaussian noise and the generated Mel-spectrogram across diverse diffusion and flow matching TTS architectures, establishing a fundamental property for watermarking.
• Robust audio watermarking can be achieved without retraining the underlying generative model, offering a "model-free" and training-free solution that significantly reduces computational burden.
• The method introduces zero computational overhead during audio generation, preserving high audio fidelity while providing robust traceability, addressing the typical quality-robustness trade-off.

Summary by alphaXiv.