Speech Watermarking for Tampering Detection Using Singular Spectrum Analysis with a Psychoacoustic Model

Unauthorized modification of speech signals can lead to misinformation, invade privacy, and reduce the reliability of individuals and agencies. This paper proposes a speech tampering detection scheme that employs semi-fragile watermarking based on singular spectrum analysis (SSA) enhanced by a psychoacoustic model. The SSA technique is used to analyze speech signals and extract the singular spectra. The watermark is embedded into speech signals by modifying some parts of the singular spectra, suggested by psychoacoustic model, to maximize transparency. After embedding the watermark into the host signal, it is possible to verify if someone has tampered with the signal. To determine whether a watermarked signal has been modified, the extracted watermark from the watermarked signal is compared to the original watermark. The evaluation results demonstrate that the proposed scheme is sensitive to malicious attacks, effectively detecting any deliberate attempts to modify the speech signal. However, it remains resilient to unintentional alterations, such as those caused by transmission errors or signal processing operations. Additionally, the evaluation results highlight the effectiveness of the proposed scheme in preserving the quality of the watermarked speech signal. The embedded watermark introduces minimal distortion, ensuring that the overall quality of the speech remains intact. Compared to a conventional SSA-based method, the proposed approach yields a notable improvement in sound quality by 7.69% and 6.91% for the objective difference grade and signal-to-distortion ratio, respectively, and the log-spectral distance is reduced by 52.17%.