Detection Of Voice Transformation Disguise Based On Deep Residual Net

Voice transformation (VT) disguise is used for de-identification which can easily deceive state-of-the-art automatic speaker verification (ASV) systems by dramatically increasing false reject rates, and thus it presents serious threats to social security. However, the researches on VT detection are insufficient. Therefore, in this paper, we propose a detection algorithm of VT disguise based on deep residual net. The proposed network structure can automatically extract deep features to distinguish disguised voices from genuine ones without gradient degradation as the number of layers increases. The experimental results show that our network can achieve accuracy rate over 96%, which outperforms the reported research efforts.