Research On Voiceprint Attack Detection Method

Voiceprint attack detection technology is a hot spot that needs great attention after the rapid development of voiceprint recognition.Nowadays,voiceprint recognition is more and more widely used in various fields,related research shows that voiceprint recognition technology is very easy to be attacked by malicious fraud,mainly divided into synthetic attack and playback attack.As the attack means become more convenient and effective,the application and promotion of voiceprint recognition system is facing a great threat.Therefore,it is of great significance to study the system of detecting synthetic and playback speech attacks.This paper mainly focuses on two aspects of synthetic speech attack and playback speech attack.It is mainly reflected in the following three aspects: firstly,the speech features are explored,and the new features that can distinguish attack speech and real speech are proposed,which are mainly divided into two categories.One is based on residual signal features,which are residual linear frequency cepstrum coefficient(RLFCC),residual constant Q cepstrum coefficient(RCQCC)and residual logarithmic power spectrum(RLog Spec);the other is based on residual signal features,the new feature of harmonic plus noise model is harmonic noise subband energy ratio(HNSER).Then,the detection system based on Gaussian Mixture Model(GMM)and light convolutional neural network(LCNN)is constructed in two scenarios of synthetic speech attack and playback speech attack respectively.Finally,on the basis of the previous research work,a hybrid attack speech detection algorithm is proposed.The research methods proposed in this paper are analyzed and tested on the task and data set of the international automatic speaker verification spoofing attack assessment(ASVSPOOF2019)in 2019.Firstly,this paper analyzes the importance of the two new robust features,and the results show that the new features have good discrimination.The experimental results show that the proposed detection systems based on RLFCC-GMM,RCQCC-GMM and RLog Spec-LCNN reduce the equal error probability(EER)by about 20%,10% and 52% respectively compared with the official baseline system,which shows that the new features proposed in this paper have certain advantages in detecting synthetic speech and playback attack speech,and can well distinguish attack speech and real speech.Although the detection system based on HNSER-GMM has no lower EER than the official baseline system,considering the complementarity between features,on the basis of the previous single system experiment,the EER of different systems is reduced by more than 30% compared with the baseline system through score fusion technology,which shows that there is information complementarity between different systems.Finally,on the basis of the previous experiments,the problems of hybrid attack are analyzed.On the premise of unknown attack mode,a hybrid attack detection system is constructed from the perspective of multi-system score fusion.The results show that the performance of the hybrid attack detection system is improved by about 90% compared with the baseline system.To sum up,the research on new features and hybrid attack detection system in this paper provides a theoretical reference and practical guidance for the development of voiceprint attack detection technology in the future.