Desigh And Implementation Of Anti-spoofing Voiceprint Recognition Based On Deep Learning

Information technology have developed rapidly and resulted in the vigorous growth of personal data,access control and privacy protection of information have always occupied the focus of people's attention.Due to the difference in physiological structure and personal vocalization way,everyone has a unique voice.And as speech is the most commonly used method in our daily communication,using one's voice to verify his identity is safe and convenient,and is easy to popularize on a large scale.In many information security fields,such as electronic payment,smart door lock and guard security,voiceprint recognition has gradually become the most mainstream trend of identity authentication method.Although researches related to voiceprint recognition have made great progress,but there are still some unsolved problems before putting it into real application.As the actual application scenes of voiceprint recognition system may be in pretty complex and diverse environments and system may face various types of unknown spoofing voices.Main spoofing voices are generated by text to speech systems,speech conversion systems and replay voices.Traditional voiceprint systems lack the detection function of spoofing voice inputs,and will greatly risk system performance.With the concern of above problem,this paper proposes a voiceprint recognition algorithm with anti-spoofing attack function based on deep learning.The proposed algorithm extracts multiple spectrogram features with different window settings as inputs,different resolution spectrograms are conducive to extract more efficient and global information from original speech.Based on deep residual network and channel attention features fusion mechanism,high-level speech representations are allowed to extracted,and then detection of spoofing speech and speaker verification are realized.The proposed algorithm conducted series of comparative experiments on ASVSpoof 2019 dataset,results proved that multiple spectrogram features with deep residual model proposed in this paper can effectively detect multiple unknown spoofing speeches.Using proposed algorithm,this paper also designs and implements a voiceprint recognition system with anti-spoofing capability,and the design and functions of each component module in the anti-spoofing voiceprint recognition system are explained in detail.