Audio watermarks are designed to be nearly imperceptible to human listeners. Yet they can still perturb the data distribution observed by anti-spoofing models—a form of domain shift that may degrade detection performance. This page is a companion demo for our paper, The Impact of Audio Watermarking on Audio Anti-Spoofing Countermeasures.
This page provides a simple A/B showcase. Each row pairs an original utterance with its watermarked counterpart. Click Show figure to inspect a compact visualization: (1) waveform overlay, (2) waveform difference (WM − ORIG), (3) log-Mel of the original, and (4) log-Mel difference in dB.
Below we highlight a slice of the Watermark-Spoof (Seen) dataset — comparing LA21 (watermarked) against its LA21 (original) counterpart. Even when the audio sounds similar, these figures reveal non-zero, structured differences (e.g., narrow-band boosts, wide-band noise-like perturbations, or echo-like periodic patterns), which are precisely the kinds of shifts that can affect spoofing countermeasures.
For clarity, labels are grouped as DNN watermarks (e.g., timbre, wavmark, audioseal) and handcrafted watermarks (all others). This demo shows only a subset — currently timbre and BlindsvdMethod.
| Watermark (label) & Class | UTT | Original | Watermarked | Controls |
|---|
python -m http.server). file:// may block resource loading.