Research On Voiceprint Recognize Algorithm Based On Convolution Neural Network

As an important studying orientation in the voice recognize area,technology of voiceprint recognize keeps an essential role for the voice scientists.Voiceprint recognize can occupy outstanding applying stages in various fields like finance,law,and national defense.Voiceprint shares the same biometric with face and fingerprint.Therefore,it can tell the identity of the speaker.As the hardware basement for voiceprint picking,microphone becomes quite common with the widely use of the smartphone.Plenty of user-voice data also become available.Under this circumstance,the original voiceprint recognize system which based on probability scoring model can hardly work well due to the low running speed.But continuous developing of deep-learning and neural network makes the applying of voiceprint recognize which is suitable for data processing come true.The main point of this paper is to come up with the application of convolutional neural network to the voiceprint recognize technology,which is based on the study of voiceprint recognize algorithm.Original voiceprint recognize technology applies Mel-Frequency Cepstral Coefficient,also called MFCC as the representation parameter.This paper analyst the principle of human vocal and auditory features,thereby upgrade the related parameter.Extract the MFCC parameter from the mixed Mel-filter bank after the structure optimization based on the frequency distribution of voiceprint characters.At the same time,by integration of the pitch frequency based on original voiceprint characters,gain new feature parameters.Applied Gaussian Mixture Model-Universal Background Model,which called GMM-UBM,to organize contrast experiments,and verify the validation of feature parameters upgrading.This paper upgrades the convolutional neural network structure based on the human auditory features and at the same time,cancels the over-all sharing function of the convolution kernel based on AlexNet classical convolutional neural network structure.Due to the different distribution of voiceprint features among different frequency ranges,pick independent convolution kernels to ensure the most obvious extraction of voiceprint features.In the network training area,extract spectrogram from every voice samples in the database as the network input feature.Organized two test groups to compare the different performance between the new model and the other one which is based on MFCC-GMM-UBM.The result tells that the convolutional neural network structure-based voiceprint recognize technology shows better performance.As plus,the studying method has some reference value in the framework of voiceprint recognize system under the big-data background.