VoxWatermark: A Large-Scale Benchmark for Audio Watermark Detection under Perturbations

2026-06-13

AI Summary

This work addresses the lack of a unified benchmark for evaluating the robustness and detectability of audio watermarking methods under realistic distribution shifts. To this end, we propose VoxWatermark—the first large-scale audio watermarking benchmark encompassing four neural and six traditional watermarking algorithms, multilingual speech corpora, and perturbation models spanning no-box, black-box, and white-box settings. Building upon this benchmark, we further introduce AudioWMD, a deep detection framework designed for cross-distribution and multi-method scenarios. Extensive experiments demonstrate that AudioWMD achieves superior robustness and scalability under complex perturbations and distribution shifts. The dataset and code are publicly released to facilitate future research.

Abstract

With the rapid deployment of speech generation systems in open environments, providing verifiable source attribution and copyright accountability for audio content has become critical. A gap in current research is the lack of a unified benchmark that systematically compares different watermark injection methods under realistic distribution shifts. To address this, we build VoxWatermark by applying 10 watermarking methods (4 neural and 6 traditional) with unified injection and annotation on multilingual, multi-source corpora, and introducing no-box, black-box, and white-box perturbations to simulate real recording and transmission conditions. Based on this benchmark, we propose AudioWMD as a robust baseline detector for large-scale, multi-method, cross-distribution settings. Results show that injection-method diversity and distribution shifts affect detection stability, while validating the effectiveness and scalability of AudioWMD. Dataset and code are publicly available.

Problem

• audio watermarking
• benchmark
• distribution shift
• watermark detection
• speech generation

Innovation

• audio watermarking
• benchmark
• distribution shift
• robust detection
• multilingual corpus