Audio Watermarking Demo
Supplementary Material (Anonymous)
Abstract
Modern audio content is often created by mixing stems from multiple sources. This work studies separation-first multi-stream audio watermarking, where each stem carries an independent watermark. We demonstrate that naive combinations of watermarking and source separation lead to poor post-separation decoding accuracy. By jointly training the watermarking system with the separator, watermark robustness after separation can be significantly improved while maintaining perceptual audio quality.
Workflow Overview
The following diagram shows the separation-first watermarking pipeline.
Audio Demonstration: Baseline, Joint, Finetuned
The examples below demonstrate the separation-first watermarking pipeline shown above. In each case, the left column corresponds to the Accompaniment stem and the right column corresponds to the Vocal stem. Each stem is independently watermarked with a different key and bit sequence. We first show the original and watermarked stems with their embedded bits. The watermarked stems are then mixed, separated using the Demucs separator, and decoded from each separated stem. For each method (Baseline, Joint, Finetuned), we show the stem-level embedding stage, the mixed audio stage, and the post-separation decoding stage with BER. Incorrect bits are highlighted in red.