Speaker Verification And Anti-Spoofing Attacks Technology Based On Deep Learning

The speaker verification can authenticate the speaker identity by analyzing his speech sample.Compared with other biometric authentication systems,it has many advantages,including convenient collection,non-contact,high recognition efficiency and good user feedback.Hence,it is of great research and practical value.However,it encounters the spoofed attacks from the advanced speech synthesis algorithms and high fidelity replay devices,which pose great threat to the security.This paper will study the core technologies of speaker verification and anti-spoofing to optimize the generalized performance of speaker verification system.Furthermore,a multi-feature multi-task(MFMT)mechanism is proposed to improve the performance of binary decision between genuine and spoofed speech.The main work of this paper is as follows.(1)Construct the speaker verification system with x-vector using different neural network architectures,such as TDNN,LSTM and Attention,etc.In order to improve the generalized performance of systems on open sets,this paper proposes the Filler universal node,which can optimize the framework and improve the recognition performance.(2)A multi-feature integration architecture is proposed to improve the security performance of anti-spoofing attacks for speaker verification system.The stitching layer,embedded within a single network,integrates the nonlinear representations from different acoustic features to obtain the high-level representation.This method can mine different time-frequency resolution at different scaled-based features.(3)Considering different kinds and unknown spoofed attacks,this paper proposes to use the multi-task learning(MTL)to extra classify the different spoofed attacks for improving the generalized performance of the binary decision task.Meanwhile,in order to optimize effectively the propagation of shared representation,a butterfly-unit(BU)is proposed to adjust the influence of gradient descent for different tasksIn this paper,the generalized performance of the Filler universal node is verified on NIST SRE10 and SRE18,and the effectiveness of the multi-feature multi-task mechanism is verified on ASVspoof2017 and 2019.In the public submissions of evaluations,we achieved the 13th out of 47 world teams on SRE18 CMN2 and the 7th out of 52 world teams on ASVspoof2019 PA,respectively.