Research On Voiceprint Recognition Based On Non-target GMM And Neural Network

Voiceprint recognition is also called speaker recognition,which is a kind of biometric authentication technology.This is to identify the speaker's identity based on the personal feature components contained in the voice,so it is also called speaker recognition.The theoretical basis is that each person's voice is different.This difference is manifested in many ways.It can be the difference in physiological physical structure,and it can be the habit and control of behavioral speech.In actual life,people mostly recognize the speaker in terms of timbre,tone,and speed of speech.The technical principle is to extract the speaker's speech features,which are used to store or compare with the stored speech features and give the judgment result.Voiceprint applications are increasingly widespread,such as criminal investigation,security,software security and other fields.The development of voiceprint recognition technology to the present,there are mainly two categories of text-related and text-independent in classification.Text-related research is relatively mature in application,but text-related features greatly limit its scope of application.Now the research focus is more on text irrelevance.Related methods include Gaussian model GMM,universal background model UBM,FA factor analysis,JFA joint factor analysis,and full factor space I-vector.The evolution of the above methods is basically based on the original Gaussian model GMM.On the basis of GMM,on the one hand,the channel robustness and accuracy are developed,on the other hand,a vector model based on GMM parameters is constructed to classify and characterize voiceprint information.In this paper,we study the dynamic regularization method of DTW applied in voiceprint recognition.On the premise of not reducing the accuracy of the amplitude,the redundant data of some trainings will be reduced to reduce the amount of data and calculations in the process of fitting and testing,in order to achieve Recognition is achieved in an embedded system with limited resources.At the same time,the posterior probability of the non-target speaker GMM to the target speaker is applied in the neural network.The experimental results show that the recognition accuracy rate of the target user GMM model is 88.73%,while the recognition accuracy rate of the GMM-BP model reaches 92.45%.The tasks of this article include:(1)Introduces the development history and current research status of voiceprint recognition,including the task of voiceprint recognition,evaluation criteria,related public data sets,andrelated algorithms.(2)Investigate a large number of documents,and deeply analyze and introduce the voiceprint recognition algorithm and its inheritance relationship.(3)In-depth exploration of Gaussian mixture model and the application of BP neural network.