Research On The Voiceprint Recognition Algorithm Based On Improved Time Delayed Neural Network In Noisy Environment

At present,deep neural network is used by more and more researchers in voiceprint recognition technology,especially time-delay neural network(TDNN).When time-delay neural network is applied to voiceprint recognition,there will be some limitations with the increase of network depth.In this dissertation,the shortcomings of time-delay neural network are improved,and an improved time-delay neural network with noise pattern recognition algorithm is proposed.First of all,as the depth of TDNN increases,the accuracy will gradually reach saturation,which makes the network performance degrade rapidly.However,this phenomenon is not caused by over-fitting,because the error increases with the increase of the number of training layers.The error generated during training leads to the occurrence of network degradation,which makes the network training ineffective.In this dissertation,the input layer of the residual neural network can modify the output layer to reduce the loss of information,and a voiceprint recognition algorithm based on residual neural network is proposed.The residual short circuit connection is directly introduced into the delay neural network,and the characteristics of residual can reduce the loss are used to alleviate the degradation of the delay neural network and improve the robustness of the system.The residual neural network is directly introduced into the delay neural network structure and three experimental schemes are designed.The optimal scheme of residual short-circuit connection is found.The results show that the effect of introducing two-layer residual neural network in TDNN1 layer and TDNN4 layer is the best.The input layer of residual neural network can play a better role in the continuous learning and correction of the output layer,which ensures the integrity of information and improves the overall voiceprint recognition rate of the system.Secondly,with the increase in the number of hidden layers in the delayed neural network,it is prone to the phenomenon of gradient disappearance or gradient explosion,which makes the neural network unable to converge.Gradient disappearance or gradient explosion directly affect the recognition rate of the algorithm in noise environment.From the point of view of solving this problem,this dissertation uses batch normalization to limit the average and variance of each layer in a certain range to reduce the shift covariance,thus improving the generalization ability of the network.A batch normalization based voiceprint recognition algorithm(BN-TDNN)is proposed.Batch normalization is added to each layer of the delay neural network except the output layer.The experimental results show that compared with the baseline system data,the voiceprint recognition rate of the system based on BN-TDNN network structure is improved.Finally,the above two algorithms are combined to design a voiceprint recognition algorithm based on improved time-delay neural network in noisy environment,that is,the residual neural network and batch normalization are directly introduced into the time-delay neural network,and the improved time-delay neural network(ResTDNN)voiceprint recognition system with noise is tested and analyzed.The experimental results on the Aishell-1 dataset show that,compared with the TDNN algorithm,the equal error rate(EER)of the ResTDNN algorithm in the noise-free environment and the noise environment is reduced by 30.4 %and 35.8 %,respectively.The experimental results confirm that the system performance of the ResTDNN algorithm is more effective than that of the residual or batch normalization used alone in the delayed neural network system,and the voiceprint recognition rate is also higher than that of the TDNN algorithm,which verifies the effectiveness of the ResTDNN network for voiceprint recognition.