VoxWatermark benchmark tests 10 audio watermark methods against real-world distortions

With the rapid deployment of speech generation systems, verifying source attribution and copyright accountability for audio content has become critical. Researchers have built VoxWatermark, a large-scale benchmark that systematically compares 10 watermark injection methods — 4 neural and 6 traditional — under realistic distribution shifts. The benchmark uses multilingual, multi-source corpora and applies no-box, black-box, and white-box perturbations to simulate real recording and transmission conditions. This addresses the gap of lacking a unified tool for comparing watermark techniques fairly.

As part of the work, the team proposes AudioWMD, a robust baseline detector designed for large-scale, multi-method, cross-distribution settings. Results demonstrate that injection-method diversity and distribution shifts affect detection stability, but AudioWMD proves effective and scalable. Accepted at Interspeech 2026, the dataset and code are publicly available, offering a foundational resource for researchers and practitioners in audio security and AI-generated speech accountability.