The Research Of Cross-Domain Voiceprint Recognition Algorithm Based On Adversarial Learning

Voiceprint recognition is an important biometric identification technology.Since it has the advantages of convenient collection,high acceptance,and protection of user privacy,it has broad research and application value in various fields.However,due to the complex acoustic environment in reality,voiceprint recognition faces many challenges such as noise interference,different speaking styles,and multiple speakers.Among them,the distribution difference between the training domain and the test domain will lead to a significant reduction in the model effect during testing,that is,a cross-domain problem,which hinders the large-scale application of voiceprint recognition technology.In response to this problem,on the basis of the research status of cross-domain problems of voiceprint recognition at home and abroad and the investigation of related technologies,based on the method of adversarial learning,this thesis studies the voiceprint recognition technology in cross-domain situations.The main research work of this thesis can be summarized as follows.The first research is voiceprint recognition method for domain adaption based on masked recurrent generative adversarial network model.Aiming at the domain adaptation problem that the model does not perform well on the test data when there is only a single source domain dataset and a single target domain dataset,based on the generative adversarial network,this thesis proposes a masked recurrent generative adversarial network model,which enables it to map target domain information to the source domain on the premise of retaining the speaker identity information to improve the effect of domain adaptation voiceprint recognition.The network model builds two pairs of generator and discriminator with completely symmetrical structures,and gradually enhances the generator's domain conversion ability by designing adversarial and cycle adversarial losses.In addition,the retention ability of speaker identity information and domain conversion ability can be increased by designing cycle consistency loss and identity loss.Also,the ability of the generator to extract feature context information is enhanced by designing a mask processing module.By completing the experimental verification on multiple voiceprint recognition databases,it is proved that the method can effectively alleviate the domain adaptation problem.The second research is voiceprint recognition method for domain generalization based on adversarial domain separation autoencoder model.Aiming at the problem of domain generalization that the model does not perform well on the test data in the case of multiple types of source domain databases and multiple types of target domain databases that have not participated in training,this thesis proposes an adversarial domain separation autoencoder model.It cleanly separates speaker identity information and domain information,so as to obtain domain-invariant features in source and target domain data,that is,pure speaker identity information,to improve the effect of domain generalization voiceprint recognition.The network model enables the network to adapt to multi-source inputs by taking the source domain labels and features together as input and enhancing the discriminator's ability to distinguish domain types.Two encoders are designed to extract the speaker information and domain of the same input respectively so as to ensure the validity of domain separation by means of cross-reconstruction.By establishing content reconstruction loss and domain reconstruction loss,the integrity of information after domain separation is ensured.The proposed method is experimentally verified on multiple voiceprint recognition databases,and the effectiveness of the algorithm in alleviating the problem of domain generalization is proved.