Modern audio watermarking systems often rely on learned encoder-decoder models that achieve strong robustness, but their behavior under new or unseen processing is typically established through empirical testing rather than theoretical guarantees. A new paper on arXiv introduces CochlearMark, a training-free alternative that aims to address this limitation.

CochlearMark is described as an audio watermark that uses phase-locked auditory nerve representations. By drawing on principles from auditory neuroscience, the method embeds watermarks without requiring a trained neural network, potentially offering a more principled and generalizable approach to robustness.

The paper's abstract does not provide details on the embedding or extraction process, nor does it present experimental results. It positions CochlearMark as a departure from the dominant learned-model paradigm, but the full methodology and evaluation remain to be seen in the complete manuscript.