| With the continuous development of artificial intelligence,more and more attention is being paid to the security of biometric identification systems,and speaker identification has gained widespread attention in areas such as national security,justice,remote authentication for banking and insurance,and security access control.How to accurately and quickly identify a person and maintain the privacy of the user's information has become an urgent problem to be solved.Earlier use of digital passwords and dynamic commands risks leakage and forgetfulness,while speaker identification,portable and convenient,spares the users from having to remember passwords or carry extra command-receiving device.However,speaker recognition systems are subject to a variety of attacks.For example,artificial intelligence is used to convert synthetic speech and record playback of speech to attack a particular speaker recognition system.This means there is still huge challenge facing the widespread use of speaker recognition system.Therefore,it is highly necessary and of great realistic importance for further investigation into speaker recognition systems with attack-detecting function,which is just the very basis and focus of this paper.Inspired by the ASVspoof 2017 and ASVspoof 2019 challenge,this paper investigates the algorithm of the speaker recognition authentication system on the basis of the current state of research and practical needs home and abroad.A speaker recognition authentication system that detects speech attacks is obtained by combining speech attack detection system with classical Gaussian mixture-general background model and deep neural network-based speaker recognition system on the framework of the speaker recognition system.The speech features are investigated in depth for the problem of robustness of the speech detection system.The major content of this paper is as follows:(1)Extraction and analysis of acoustic features.In this paper the difference between real speech and attacking speech is analysed,by pre-processing speech signals to enhance acoustic features.(2)A speech attack detection system based on traditional recognition models is constructed.The classical classification model algorithm is mainly studied,and a Gaussian mixture model is built to classify the speech,and the model is optimised by an adaptive algorithm and iterated using the maximum expectation algorithm to determine the parameters.Finally,the experiment results are compared and analysed.(3)Construction of a speech attack detection system for speaker recognition based on residual networks.The speech is processed such as Fourier transform to obtain a speech spectrogram,and extended or cropped to obtain uniform features.The Equal Error Rate(EER)and Tandem Detection Cost Function(t-DCF)of the Spec-Res Net model are significantly reduced.(4)To construct an end-to-end high-resolution complementary spectrogram speech spoof detection system,phase spectrogram is used as complementary information to the amplitude spectrogram,and the probability density spectrogram is introduced to combine the three into a complementary spectrogram feature with high resolution,and the end-to-end network structure is obtained by improving the residual neural network,and the frame-level speech features are converged into speech-level features by gated recurrent units.The performance of the amplitude spectrum,phase spectrum,probability density spectrum and complementary spectrum are compared and analysed.The EER and t-DCF of the optimal speech attack detection system on ASVspoof2019 physical access are 1.75 and 0.0491,respectively,which are 72% and 50% lower than the EER and t-DCF of the Spec-Res Net attack detection... |
PDF Full Text Request