Research On Detection Algorithm Of Speech Spoofing And Its System Implementation

Speech spoofing refers to the production of artificial speech by certain disguising methods for deception.Existing studies have shown that speech spoofing has the ability to deceive the Automatic Speaker Verification(ASV)system,and thus presents damages and threats to social security.Therefore,the research on speech spoofing detection is of significant theoretical and practical value.The existing research efforts mainly employs traditional machine learning methods or general convolutional neural network for detection.Traditional methods relies on manual designed features,which is difficult to ensure robustness,and some deep-level features may be lost in preprocessing as well.The general graph convolutional neural network has good classification accuracy,which is vulnerable to unknown spoofing attacks.Hence,in this thesis,we investigate speech spoofing detection based on graph attention network,and develop a frame work for graph data.Both node information and the structural information are considered for graph data,resulting that graph attention neural network can not only learn the node features automatically,but also learn the implicit association information between the nodes in the graph,predict unknown spoofing attacks.The main contributions of this thesis is as follows:1.We propose a spoofing speech detection algorithm based on graph attention network model.The proposed network model is optimized from the ordinary graph attention network.Original audio temporal waveform is treated as the input into the network to avoids feature loss.Residual blocks are added,and then two graph attention networks are established based on time domain and frequency domain to extract deep features.The experimental results show that the detection algorithm achieves an equal error rate(EER)of 1.85% on the LA logical access corpus of ASVspoof2019.2.In order to facilitate the visualization of the model and the persistence of experimental data,we design and develop a general spoofing detection model framework,and employ the Dense Net as the underlying model to build the detection system.The basic functions of this framework,such as the adjustment of hyperparameters in the visual interface,the display of training data of the algorithm model,the persistence of experimental data,and the one-key audio preprocessing operation,have laid the foundation for the subsequent theoretical research and development of the laboratory.For summary,the proposed detection algorithm can be used as a detection module in ASV systems for robustness enhancement,which is of great significance to the construction of information security.The implementation of the general framework provides support for the development of the subsequent research on speech spoofing detection.