Anti-counterfeiting noise addition strategy model for voice cloning

This study aims to explore effective anti-counterfeiting noise addition strategies for voice cloning technology based on VITS. Initially, multiple public voice datasets were collected and normalized to ensure data consistency and comparability. Subsequently, various noise addition processes were applied to the original voice samples based on the proposed anti-counterfeiting methods. In the experiments, we systematically compared the effects of different noise addition methods, including Gaussian noise, salt-and-pepper noise, blank frame combination operations, and frame window standard convolution, on the voice cloning model. The evaluation was conducted using MCD (Mel Cepstral Distortion) scores for voice synthesis quality and MOS (Mean Opinion Score) to weight the merits of each method in terms of anti-counterfeiting effectiveness. The conclusion drawn is that a combination of frame window standard convolution and Gaussian noise can most effectively interfere with AI cloning models without affecting the transmission of vocal information.