Research On Multi-parameter Channel Information Playback Attack Detection

Biometrics has broad research prospects.As an important part of biometrics,speaker recognition is involved in many aspects of people's daily life.With the popularity of high fidelity recording equipment and playback equipment,the security of speaker recognition system is facing a serious challenge of recording playback attack.Because the playback attack voice has the same voiceprint as the real voice,it is difficult for the conventional speaker recognition system to effectively identify the authenticity of the voice,and there is noise in the environment,which will interfere with the recognition of the system to a certain extent,which is also a challenge of the robustness of voiceprint recognition system against recording playback attack is required.In this thesis,a detection algorithm of recording playback attack based on channel information is proposed.Legendre coefficient and its statistical characteristics are used as the main discrimination basis.At the same time,speech fundamental frequency feature and MFCC feature are used as the auxiliary discrimination index.A decision fusion strategy based on support vector machine is used for discrimination.Different weights are given to the features,and the decision results are obtained respectively by the decision fusion formula calculates the overall probability and compares it with the acceptance threshold as the final decision result.This thesis mainly includes the following contents:(1)This thesis introduces the development of voiceprint recognition system,and focuses on the challenge of recording and playback attack faced by the system,and then makes a detailed introduction to the research status of the challenge at home and abroad.(2)The common processing methods of speech signal processing are described in detail,and then the popular feature parameters and extraction steps in the current speech signal processing are discussed,and the existing recording playback detection algorithm is analyzed and introduced.(3)On the basis of the existing research,we further improve the system model,introduce multiple parameters,build a multi parameter channel information playback detection model,and discuss the overall framework of the model,related parameters such as fundamental frequency characteristics and channel mode noise statistical characteristics,support vector machine in modeling means and the advantages of decision fusion strategy.(4)Aiming at the improved multi parameter channel information playback detection algorithm in this paper,the weight parameter selection analysis experiment,the system performance experiment under different SNR and the system robustness experiment when different recordings in the same data set are used as attack test speech are carried out.The experimental results show that,compared with other existing methods,the combination of multiple features can effectively detect playback speech attacks and improve the robustness of the system.The average recognition rate is improved by 1.5% in noisy environment.