Speech Spoofing Detection Based On Residual Network

Speech spoofing can be divided into two categories according to different purposes: 1)impersonation spoofing,which aims to impersonate a target person by means of speech synthesis,voice transformation or re-capture.2)concealment spoofing,also known as voice transform(VT),which aims to hide speaker's identity by changing the voice without impersonating a target,The existed researches have revealed that speech spoofing can deceive speaker recognition systems,which presents a serious threat to social security.Therefore,detection of speech spoofing is of significance of research and practice.Hence,in this thesis we research on the detection algorithms of speech spoofing.The main contributions of this thesis are as follows.1.A speech spoofing detection algorithms which employs discrete cosine transform and a Resnet with deep feature extraction capability for classification is proposed.The experimental results show that an accuracy rate over 99.9% on the validation set and an EER of 2.32% on the evaluation set are achieved,indicating that it has good detection performance.2.A detection algorithm of voice transformation spoofing using multi-resolution DCT as input features and a residual network with channel fusion technique is proposed.The detection accuracy achieves over 97% in the case of intra-database.The experimental results indicate that the multi-resolution DCT features and the proposed network structure can better characterize detailed information of the frequency distribution of speech signal,and and thus achieve a good detection performance.In this thesis,we propose detection algorithms for different types of speech spoofing,which can be integrated into automatic speaker recognition system to enhance robustness,and is of great significance for social security.