Speech Spoofing Detection Based On Dense Neural Network

Reported efforts have revealed that speech spoofing,including voice transformation(VT),voice conversion(VC),speech synthesis(SS)and replay attack,can effectively deceive current ASR systems and present serious threats to social security.Hence,it is of great theoretical and practical significance to study the detection of speech spoofing.The existing researches mainly employ traditional machine learning methods or conventional convolution neural network methods,in which the extracted features may contain insufficient information or depth.Therefore in this thesis,we study speech spoofing detection based on dense neural network,which can automatically extract deeper and better features to improve detection performance.The main contributions of this thesis are as follows.1.A VT detection algorithm based on convolutional neural network(CNN)model is proposed.the proposed network model structure is optimized based on dense neural networks in convolutional neural networks.This network model consists of 167 network layers to extract deeper features than traditional network,and to further obtain higher detection accuracy.The experimental results show that the accuracy rates in the intra-database tests are over 97.4%,and the accuracy in the inter-database tests are also over 91%,indicating that the proposed algorithm has good robustness.2.For the detection of VC and SS,the deception detection model constructed in this paper mainly consists of three dense blocks.On the basis of VT detection,the deeper network layer is used,with a total of 175 network layers.The convolution kernel and pooling kernel with special structures are adopted,which can pool the spectral features along time axis and prevent over-fitting.The experimental results show that the accuracy rates are over 97% the intra-database tests,and the accuracy rates are over 94% on the inter-database tests,and the Equal Error Rate(EER)achieves 1.44%,indicating that the proposed model has a good performance on VC and SS detection.The speech spoofing detection algorithm based on dense neural network proposed in this thesis can be used as a detection module based on Automatic Speaker Recognition(ASR)system,which makes it have the ability to resist speech spoofing attack and has important significance for information security construction.