Research On Key Approaches Of Replay Detection Based On Deep Learning

With the development of Automatic Speaker Verification(ASV)technology,voiceprint have begun to apply in various fields of society,including but not limited to voiceprint payment,voiceprint unlocking,and home voiceprint control.However,the security of ASV has also been questioned.In order to prevent imposters from attacking the system through various means,ASV anti-spoofing technology has emerged.Especially in recent years,inspired by the ASVspoof Challenge,the academia has paid more attention to ASV spoofing attack detection,and many excellent models and approaches have been built.Based on the international standard asvspoof2019 data set,we build 1.A CQCC?GMM baseline detection system.2.Logspec RESNET detection system combining four kinds of residual networks with logspec(log spectrum)acoustic characteristics.3.A variety of ECA based on attention mechanism and shallow residual network?Resnet18is the representative of the detection system.4.RF based on score domain fusion?Att?RESNET detection system and FAF based on feature domain fusion?att?Fusion?RESNET detection system step by step.Through experimental analysis and comparison,it is concluded that the shallow residual networks resnet18 and resnet34 have better detection performance.Then we choose the two to continue to optimize and introduce the attention mechanism into the network.We use four methods,namely se module,ECA module,CBAM module and FT CBAM module,to form a new model combined with resnet18 and resnet34.The experimental results show that each model has its own advantages and disadvantages in parameters,training speed and test results?Resnet18 model achieves the best detection effect with EER(equal error rate)of 0.688%.Finally,the four models are fused based on score domain and feature level,and the best fusion model FAF is obtained?att?Fusion?RESNET,EER was 0.612%,which was 93.79% higher than the baseline system.In addition,this paper also introduces the self-made Chinese children's voice replay attack data set--childspoof,which is the first to add a large number of children's voice to improve the diversity of the data set voice.The data set is used to test the optimal single model ECA?The performance of resnet18 proves the effectiveness and robustness of the model.In the future,it can be transplanted to voiceprint recognition system to help improve the security and reliability of the system,and contribute to the development of voiceprint recognition applications.