Research And Implementation Of Speaker Recognition Method For Anti-playback Fake Speech

Speaker recognition technology is a technology that using the features in the speech signal to identify the speaker.In the field of biometric technology,speaker recognition technology has a wide range of applications in daily security,financial security,forensic identification,and intelligence interception due to its high security.However,with the popularization of highfidelity audio equipment,the offensiveness of replaying fake speech has increased,resulting in a decrease in the security of the speaker recognition system.In order to improve the ability of speaker recognition system to resist replaying speech attacks,replaying speech identification methods have become one of the research hotspots in the field of speaker recognition.This paper studies the speaker recognition method based on I-vector,and uses feature combination to improve the robustness and accuracy of the system.For speaker verification,this paper studies the calculation method of adaptive verification threshold on the basis of conventional confirmation threshold.In order to make the system have anti-replay voice function,the replay voice identification algorithm based on wavelet packet decomposition and convolutional neural network is studied.The main work content is as follows:(1)On the basis of speaker recognition based on I-vector,the MFCC and PNCC feature parameters are combined for the problem of insufficient single feature information,and the combined feature parameters are optimized by the feature selection method based on genetic algorithm.Experiments show that the feature combination method can improve the accuracy of speaker recognition.(2)For the speaker verification system in the open set situation,the conventional speaker verification threshold is introduced.Aiming at the problem of insufficient adaptation of the conventional threshold in practical applications,the calculation method of the adaptive verification threshold is studied.Aiming at the problem of abnormal values in speaker scores,the abnormal score elimination algorithm is studied,and the adaptive verification threshold calculation method is combined to calculate the verification threshold.Experiments show that the adaptive verification threshold calculated after eliminating the abnormal score can effectively improve the performance of the speaker verification system.(3)Aiming at the playback voice attack problem,the difference between real voice and playback voice is analyzed,and the playback voice identification algorithm based on wavelet packet decomposition is studied,and the effectiveness of the algorithm is verified through experiments.In order to improve the generalization performance of the algorithm,the playback speech identification algorithm based on convolutional neural network is also studied.Experiments show that the generalization performance of the algorithm has been improved.(4)Development of a speaker recognition system that resists playback speech attacks.The overall design of the system is introduced,the functions and implementation methods of the main modules are discussed,and the performance of the system is verified by experiments.