Research On Voiceprint Recognition Based On Margin Adaptive Method And Relational Network

In recent years,the rapid development of science and technology has brought great convenience to people's social life,and has also put forward higher and higher requirements for the security of personal identity authentication.At present,identity authentication methods based on credentials or passwords can no longer fully meet people's needs,at the same time,biometric-based authentication technologies have attracted more and more attention from society and researchers due to their security and reliability.Speech is the most direct and convenient way for people to communicate in daily life.Therefore,speech-based identification technology has also become a research focus.This technology is called voiceprint recognition technology.Recently,a series of breakthroughs made by the application of deep learning technology in the field of voiceprint recognition,especially the proposition of a loss function based on the margin,has greatly improved the performance of the current voiceprint recognition system.However,the current voiceprint recognition technology still has definitely room for improvement.For example,for the loss function based on margin,how to set the margin is a difficult point.If the margin is too small,the features learned by the network are not discriminative enough.On the other hand,if the margin is too large,then the network will fall into non-convergence state.There is no perfect solution to balance between these two sides until now.In addition,after using the neural network to extract the voiceprint features,we need to measure the similarity between them.The current measurement method is mainly cosine similarity,and cosine similarity is an artificial,non-parametric similarity measurement method,which lacks a deeper modeling ability for voiceprint similarity.Therefore,this thesis intends to deal with the above two issues.(1)For the difficulty of setting the margin in the margin-based loss function,we propose a margin adaptive algorithm in this thesis.This algorithm adapts reinforcement learning method to enable the network to continuously adjust the margin according to the training state of the network during the training process.Moreover,the margin setting of the algorithm is class-dependent,allowing the different class to choose the different margins.(2)When measuring the distance of voiceprint features,a distance learning algorithm based on deep learning,rather than traditional cosine distance,is proposed in this thesis.In this algorithm,mechanism of relational networks,which allows the network to learn and characterize similarities between features autonomously,is introduced.Experiment results show that this algorithm can make further improvements on the performance of our voiceprint recognition system.Based on the above design,a voiceprint recognition system based on margin adaptive and relational networks is built in this thesis,and its performance is evaluated on the data set voxceleb2.Experimental results show that our method has a relative reduction of 14.4% in equal error rate compared with the best performance method in the current academic circles,which proves the effectiveness of our method.