Research And Implementation Of Anti-spoofing Voiceprint System Based On Android

In the era of rapid development of smart phones and smart terminals,it is increasingly necessary to strengthen the protection of personal information.Voiceprint algorithm is a security verification method constructed by using human voice features.Currently,there are two main problems encountered by voiceprint systems.The first problem is that when extracting identity features through neural networks,more channels and deeper network layers are often used to construct neural networks to extract features with stronger expressiveness.However,the increase in network complexity not only increases the difficulty of training the network,but also makes the system bloated when deploying the model.Therefore,this paper adopts a dense neural network(Dense Net)and a dense delay neural network(D-TDNN)combined with a feedforward neural network(FNN)as the algorithm of the voiceprint system.D-TDNN adopts bottleneck layers and dense connections,which can aggregate multi-stage information from different layers,reducing computational difficulty and having fewer parameters.Through the verification on the VOXCELEB dataset,the accuracy of D-TDNN is not much different from that of similar models,but the number of parameters is greatly reduced,and the model is lighter.The voiceprint system is often used in the field of security verification,so it is particularly important to ensure the security performance of the system.The second problem is that in actual complex and diverse application scenarios,the system often faces different types of spoofed voice threats.To solve this problem,this paper proposes a Res Ne Xt speech deception detection model fused with attention mechanism-RA-Net.Compared with using a single residual neural network(Res Net),this model uses fewer hyperparameters and can more highlight high-frequency features in speech features.RA-Net adopts the method of combining residual and grouped convolution,replaces the large convolution kernel with a set of small convolution kernels,and adopts MFM(Max Feature Map)as a new splicing method.The added attention mechanism reduces the attention to edge features by learning the information of the original features.Experiments on the ASVspoof2019 dataset show that the equal error rate(EER)of RA-Net is reduced by 4.72% and 6.23% compared with the baseline Gaussian mixture model(GMM),and the EER is reduced by0.69% compared with the residual network(Res Net).and 0.89%,demonstrating the effectiveness of the model.Both the voiceprint recognition algorithm and the deceptive speech detection model studied in this paper are trained and tested on the corresponding datasets.On the basis of the algorithm,an anti-spoofing voiceprint system based on Android platform is designed for recognition.Finally,input speech on the real machine for testing,the recognition accuracy rate of the system is 81.25%,the recognition error rate is 18.75%,and the error acceptance rate is 2.5%.