Speech watermarking scheme based on singular-spectrum analysis for tampering detection and identification

This paper proposes a speech watermarking scheme for tampering detection and identification, which is based on the singular-spectrum analysis (SSA). Tampered signals can be detected and identified by analyzing signature information embedded into the speech signals. A required property of such hidden information is that it should be fragile to any malicious signal-processing operation. In this work, we use the SSA to analyze signals in embedding and extraction processes. The signature information is embedded into those signals by modifying a less-significant part of singular spectra. The evaluation results show that the proposed scheme is fragile to several malicious attacks but robust against non-malicious signal- processing operations. The sound quality of the watermarked signals also meets the inaudibility criteria. Besides detecting and locating the tampered locations, our proposed scheme can make a prediction about types and degrees of attacks. In addition, its new embedding rule enables one to fine-tune some parameters in order to improve the overall performance further.