Audio Watermarking

Audio watermarking studies how to attach hidden or audible-to-machines information to speech, music, and synthesized audio so its source or integrity can later be checked. Methods range from post-hoc signal and spectrogram perturbations to marks embedded in neural-codec latents, semantic or token representations, and generation-time TTS or music models. Researchers test detection and payload recovery under compression, denoising, resynthesis, source separation, and other edits, with benchmarks increasingly measuring false positives and distribution shifts. The central tension is making marks survive powerful transformations and removal attacks without harming sound quality, while ensuring provenance marks do not mislead or weaken deepfake detectors.