OrthoSig-Mark: Detector-Orthogonal and Content-Bound Audio Watermarking Against Anti-Spoofing False Positives and Copy-Paste Forgery

Neural audio watermarking and passive anti-spoofing detection are increasingly deployed together to establish the provenance of speech, yet their interaction has not been examined. We show that embedding an imperceptible neural watermark into authentic speech causes passive deepfake detectors to flag 61.7–70.0% of the watermarked utterances as synthetic. Spectral sensitivity analysis attributes this mark-to-frame effect to checkerboard harmonics near the Nyquist frequency that are introduced by the final transposed convolution of the watermark generator, even though the message itself occupies only the band below 2 kHz. Moreover, fixed-bit watermarks are not bound to the acoustic content and can therefore be transplanted, superimposed, or left intact by word-order splicing. We propose OrthoSig-Mark, a 64-bit watermarking framework that addresses both problems. A differentiable spectral projection with an exact second-order zero at the Nyquist frequency, trained with a detector-orthogonal loss and a self-hash invariance loss, suppresses the artifact seen by the detector. A payload that combines a synchronization-anchored acoustic similarity hash, a truncated keyed-hash authentication token, and a Bose–Chaudhuri–Hocquenghem parity code binds each watermark to its host signal and is verified by cryptographically guided soft-decision list decoding. On up to 120 utterances, OrthoSig-Mark lowers the false positive rate of the detector used in training from 70.0% to 6.7% and that of two unseen detectors from 61.7–65.0% to the clean-speech level of 5.0%, while keeping a perceptual quality score of 4.38 and 99.2% bit accuracy, and it detects all four evaluated forgery attacks. (English version and Japanese version available)