Disguised Voice Voiceprint Recognition Algorithm Based On Deep Learning

With the development of the Internet and information technology,voiceprint recognition technology has broad application value in remote identity verification and identity authentication of security departments,and it is an important exploration direction in the field of biometric identification.But the appearance of disguised voice has a significant impact on the performance of voiceprint recognition system,so it often brings troubles and crises to our lives.Aiming at the problem of poor performance of voiceprint recognition system under disguised voice,in this thesis,two aspects of feature extraction and model establishment are studied to solve the problem that the traditional Gaussian mixture model recognition system is not effective in disguised voice test.The main contributions are as follows:From the perspective of feature extraction,in the thesis,after preprocessing the speech information,the resonance peak parameters are calculated by cepstrum method,and Gammatone Frequency Cepstrum Coefficients(GFCC)is obtained by the Gammatone filter bank,then proposing to compose the formant and GFCC and its differential coefficients as a mixed characteristic parameter,while collecting normal speech and seven different disguised voice of 49 people,and experiments were conducted on these voices.Experiments show that this method of extracting mixed features from original speech data can extract feature expressions with more speaker distinguishing characteristics.The proposed method solved the low accuracy problem of the speaker recognition system under disguised voice using the traditional technology.From the perspective of model establishment,in the thesis,based on the mixture feature,we used deep belief network(DBN)to replace the Gaussian mixture model as the acoustic model of the speaker recognition system,and the dropout strategy is introduced to suppress the over-fitting problem during DBN training,at the same time,experiment with the collected speech library.It can be seen from the experimental results that the voiceprint recognition model established by DBN can better fit the feature distribution,and the classification effect is significantly improved,thus greatly improved the recognition accuracy of systems.Simultaneously,the introduction of dropout further increased the recognition rate.