| With the advent of the information era,information technology is becoming gradually mature,and the corresponding information security receives increasing attention.Automatic speaker verification(ASV)system can identify the speaker by matching the speaker’s voice.However,today’s speech spoofing technology is becoming more and more sophisticated.Speech forgers attack ASV systems by means of speech synthesis,speech conversion and speech replay.They can easily evade the detection of ASV systems and achieve the aim of deceiving ASV systems,which brings a great impact to ASV systems.Consequently,it is essential to develop effective anti-spoofing countermeasures.In this paper,we will conduct a study on speech spoofing detection.The following work has been done:(1)To address the relatively single scale of features and the lack of abundant feature content,we propose a feature pyramid residual network(FP-ResNet)model.The FP-ResNet model with diverse feature scales is enable to carry out different degrees of feature mining,and combine the residual structure to ensure the useful information is transmitted to the deeper network.This paper conducts experiments on ASVspoof2021 LA dataset.The experimental results are t-DCF=0.2545 and EER=2.80%.Experimental results show that FP-ResNet can improve the accuracy of speech spoofing detection.(2)To address the sparse connection between the features of each layer of FPResNet model,we proposes a double-path recursive feature pyramid residual network(RFP-ResNet)model which enhances the relationship between the features of each layer of the network.Subsequently,we uses two-step training to improve the robustness of the model when training the two-path network.In this paper,experiments are conducted on the ASVspoof 2021 LA dataset.The experimental results are tDCF=0.2469 and EER=2.57%.The experimental results show that enhancing the connection between the features in each layer of the feature pyramid can improve the accuracy of speech spoofing detection. |
