Research Of Robust Speaker Verification Baesd On Deep Learning

Speaker verification is a technique of judging whether a given speech is a declared speaker identity,which has important application in user identity authentication in various occasions.The speaker verification technique has achieved incredible performance for pure speech.However,the performance of the system is degraded due to the interference of noise,which has the greatest impact in practical applications.To address the issue of performance degradation due to noise interference,this study combines speech separation tasks with speaker verification.This paper proposed a joint training framework based on deep neural network(DNN)for speech separation and speaker verification,which applies the noise-robustness feature generated by speech separation to the speaker verification network,which can significantly improve the performance of the speaker verification.The proposed method consists of two parts.Speech separation module based on convolutional recurrent neural network and the end to end speaker verification module of deep residual network structure.Then the output of middle layer of the recurrent network is used as auxiliary feature,and together with the robust Filter banks(Fbanks)feature of noise speech are fed to the speaker verification are jointly optimized.The experimental results show that this method is much better than the traditional.On this basis,in order to cope with the issue of performance degradation due to the increasing of the sequence length,this study proposes to use attention mechanisms to produce a representation vectors for input different length of sequences.Each frame outputs a vector which containing information from all previous frames,so that the final feature sequence contains more speaker information.Experimental results show that the performance of joint training framework with attention mechanism has a significant improvement in multiple cases compared with baseline systems.