Application Research And Implementation Of Voiceprint Recognition

Voiceprint recognition is a kind of biometric authentication technology.Compared with other biometric authentication technologies,it has the characteristics of simple authentication method and low equipment cost.Although deep learning has achieved great success in voiceprint recognition,its huge storage and computational costs have hindered its application in resource-constrained environments.This paper mainly studies the engineering application of voiceprint recognition with limited resources.First,for the problem of voiceprint recognition with only a small amount of registrant data available,this paper designs a One vs Rest(OvR)neural network model for identity authentication.Experimental results on a text-dependent self-built data set and text-independent AISHELL-160 data set show that the average error rate of the model in closed-set voiceprint recognition is 0.13%and 0.8625%,respectively,reaching a practical level.In open set voiceprint recognition,for text-dependent data sets,the average false rejection rate and average false acceptance rate are 0.53%and 1.77%,respectively,which still have practical value.Secondly,for the problem of voiceprint recognition that has unregistered person data available,this paper designs a recognition system based on Time Delay Neural Network(TDNN).First train a TDNN model with unregistered person data,and then use this model as a feature extractor to extract the voiceprint features of registered users,and input OvR back-end model for classification and recognition.The results on the two data sets of AISHELL-16 and AISHELL-160 show that,compared with the recognition results of only registrant data,the error rate,false rejection rate and false acceptance rate of TDNN+OvR on AISHELL-16 are respectively reduced by 0.25%,0.375%and 3.25%.In AISHELL-160,it was reduced by 0.475%,0.825%,and 3.2375%,and the recognition effect was significantly improved.Further,we conducted an experimental comparison between the TDNN+OvR and TDNN back-end using classifiers such as fully connected neural networks,K-nearest neighbors,and cosine distance.The equal error rates of the four methods on the AISHELL-160 data set were 4.4375%,5.375%,6.875%and 6.25%,indicating that the identification system combined with TDNN and OvR neural network adopted in this paper has obvious advantages.Finally,this article establishes an actual voiceprint recognition system.The entire recognition system includes recording,registration,and real-time testing functions.The user only needs to click the corresponding button on the software interface to complete the target operation.It is friendly and easy to operate.