Research On Voiceprint Recognition Technology For Anti-synthesis Voice Attack

Voiceprint recognition technology is a branch of biometric recognition and one of the hot spots in the field of artificial intelligence.It has made rapid development and wide attention in theoretical research and practical application.It has a wide range of applications in the field of information security,such as pension insurance system,identity authentication system,access control system and security.However,with the development of science and technology,the synthesis of speaker's voice becomes more convenient and accurate,which makes the synthetic voice have a great threat to the information security of voiceprint recognition system,and also makes the popularization and application of voiceprint recognition system face a great challenge.Therefore,it is urgent and significant to study the voiceprint recognition system to prevent the attack of synthetic voice.Based on the competition of asvspoof 2015,combined with the actual application demand and the research status at home and abroad,this paper studies the voiceprint recognition algorithm against synthetic speech attack.On the basis of the voiceprint recognition system based on GMM-UBM,combined with the anti-synthetic speech attack detection system,we get a fusion system that can judge whether it is synthetic speech and whether it is the target speaker.In the feature extraction,the feature fusion method of Mel frequency cepstrum coefficient(MFCC)and linear frequency cepstrum coefficient(LFCC)is introduced,and the new learning filter bank obtained from the deep neural network is applied to the traditional Gaussian mixture model system.Finally,the voiceprint recognition system against synthetic speech attack is designed and implemented.In the introduction of the classical voiceprint recognition algorithm,the traditional algorithm and the deep learning algorithm are compared from the theoretical and experimental analysis,and the advantages and disadvantages of the two are compared from the results.In order to improve the performance of the system and improve the recognition rate,this paper analyzes and experiments different feature extraction algorithms,and obtains relatively superior performance by combining different features.On this basis,the algorithm principle and workflow based on deep learning are further elaborated.In the data set of asvspoof2015,there are ten different methods to synthesize the speech,the average equal error probability(EER)of the synthetic speech attack detection system based on single MFCC feature is 4.6371%,the average equal error probability of the synthetic speech attack detection system based on the fusion of MFCC and lfcc feature is 2.5967%,and the synthesis designed by the learning filter bank obtained by deep neural network learning The average equal error probability of speech attack detection system is 1.0089%.It can be seen from the results that the synthetic speech detection systems with different characteristic parameters have strong complementary characteristics.On the other hand,the voiceprint recognition system with synthetic speech detection is significantly better than the traditional voiceprint recognition system in performance.Due to the inability to distinguish synthetic speech,many synthetic speech are recognized as the target speaker.The EER of voiceprint recognition system without synthetic speech attack detection is 13.5183%.The performance of this system is very poor,and it can not be applied in the actual scene,while the EER of voiceprint recognition system with synthetic speech attack detection is 5.1064%,which greatly improves the recognition accuracy,which is also reflected the significance of this paper is discussed.