Research And Implementation Of Voiceprint Recognition Technology Based On Neural Network Feature Mapping

The recognition technology of voice pattern has been applied in many fields such as judicial identification,military security,financial fraud prevention,etc.but in the practical application environment,various background noise which affects the voice quality makes the effect of the system in practical engineering application not meet the ideal requirements,which becomes an important factor to restrict the development of the technology of voiceprint recognition.Therefore,the thesis focuses on the improvement of recognition rate of the recognition technology in noise environment.Inspired by the research of neural network based feature mapping in speech enhancement and robust speech recognition,this paper applies neural network based feature mapping method to robust speech recognition system.In this paper,the deep neural network(DNN)is trained by using the parallel corpus data set constructed by artificial method,which aims at the minimum mean square error,so that it can learn the mapping relationship between the logarithmic power spectrum characteristics of noisy speech and the characteristics of clean speech Fbanks.The trained DNN model is used as a feature extractor in the feature extraction process of the recognition system.The experimental results show that the performance of DNN output features is higher than that of traditional robust features at low SNR,while the performance of DNN output features is lower than that of traditional robust features at high SNR.Affected by the objective function of minimum mean square error,the output features of DNN can not achieve better performance than the traditional robust features at all SNR.In order to solve the problem caused by FMDNN target function,the DNN is replaced by Wasserstein to generate the counteractive network(WGAN).WGAN measures the distance between the generated data distribution and the real data distribution by Wasserstein distance,which makes the neural network better learn the mapping relationship between the features.The experiment shows that GMM-UBM is used as the recognition model compared with the traditional robustness characteristics under the test conditions of 0d B,5d B,10 d B,15 d B and 20 d B,The recognition rate of FMWGAN-MFCC output by 13%,7.3%,6.7%,6.8% and 6.3% respectively;The recognition rate of I-vector is 5.5%,1.9%,1.2%,0.6% higher than that of the other two models.This paper designs a program of recognition of the voice pattern.The neural network model trained in this paper is applied to the realization of the recognition program.After the development of the program,the performance evaluation and test of the program are carried out in the actual environment.From the simulation experiment and the actual test,the FMWGAN-MFCC based recognition system has higher recognition rate than the traditional robust feature based recognition system in noise environment,which can solve the problem of the reduction of recognition rate in noise environment.