Application Of Playback Speech Detection Method Based On AdaBoost Algorithm In Automatic Speaker Verification

In recent years,significant progress has been made in the field of speaker recognition.Automatic Speaker Verification(ASV)has been put into use in call centers,forensics,access control systems,e-commerce,etc.due to its reliability,convenience,and low cost.Due to the popularity of playback and recording of voice devices,and the emergence of audio processing software,it is easier to edit,modify,and synthesize voice data.In particular,playback voice attacks do not require any expertise in voice signal processing,and ordinary people only need one.The device that can record and play can attack the ASV system,and the voice comes from the legal speaker.In this environment,the performance of the speaker recognition system is seriously degraded,and it is difficult to achieve satisfactory results.In order to further improve the security of the ASV system,playback voice detection is a crucial research hotspot in the speaker field.The main work and innovations of this paper focus on the following aspects:(1)For playback voice detection,it is very important to select appropriate features to reflect the relevant information carried in the voice.In this paper,the Gaussian mixture model is used as a classifier,and the performance of 9 features in playback speech detection is compared experimentally.The study compared the effects of static and dynamic features on the test results and the performance of feature generalization.The experimental results show that high-frequency features,dynamic features,and features with detailed spectral information are effective for improving the detection results.However,these features are not able to fully describe the characteristics of the playback voice,and the generalization ability is poor.(2)In order to improve the anti-recording playback attack capability of the automatic speaker confirmation system,the ASVspoof 2017 Challenge focuses on playback voice detection.Aiming at the problem of limited classification ability of a single classifier,a playback voice detection method based on AdaBoost algorithm is proposed.The method takes the constant Q cepstrum coefficient as the characteristic parameter,and the mean super vector as the input of the AdaBoost algorithm,and combines the detection results of multiple classifiers to reduce the equal error rate.The influence of relationship factor,mean supervector dimension and the number of weak classifiers on the detection results is studied.In the development set and evaluation set,the error rate of the baseline system provided by the ASVspoof 2017 Challenge organizers was 11.85% and 30%,respectively.The error rate of the test method was 4.17% and 16.81%,respectively,which decreased by 65% and 44%respectively.(3)An automatic speaker confirmation system based on I-Vector and PLDA model is constructed,and the I-Vector extraction algorithm and PLDA model are analyzed.The playback speech detection method is loaded into the front end of the I-Vector/PLDA automatic speaker confirmation system for comparison experiments.The experimental results show that the automatic speaker verification system without the playback voice detection method has the same error rate of 13.51% and 22.38%on the development set and evaluation set.The security performance of the system is very low.After adding the playback attack detection method,The error rates on the development and evaluation sets were 9.17% and 13.25%,down 32% and 41%,respectively.