Algorithm Research And Design Of Anti-playback Attack Speaker Recognition System

Speaker recognition technology is one of the hot spots of voiceprint recognitiontechnology researches, which was widely used in information security such aspension insurance systems, access control systems and remote recognition systems,etc. However, with the development of technology, the speaker’s voice recording ismore and more convenient and easy to use, which makes recording replay attacksproduced a big threat on the speaker recognition information security system.Therefore, the study of prevents replay attacks recordings speaker recognition systemis of big significance and urgent needs.Based on the analysis of existing speaker recognition algorithm, this articlestudied the speaker authentication system based on Gaussian Mixture Model andHidden Markov Models and its simulation. On the basis of the analysis of distinguishbetween the original voice and playback voice, the article extracted the features thatcan correctly express the distinction between the two voices, studied theanti-recording playback system based on the SVM, and finally designed and realizedthe whole system of speaker recognition with anti-speaker audio replay attackdetection.The main contents are as follows:(1) Analysis of the current speaker recognition processing, focused onresearching the algorithm that removed the noise of voice signal and enhanced thevoice magnitude algorithms, programming to implement the algorithms andanalyzing their effects. Achieved an extraction of MFCC vector that is a voice speechfeature, analyzed existing methods of anti-playback recording attacks, made a pointthat according to the long-term speech signal characteristics of the channel patternnoise to distinguish the two voices.(2) In the study of the basic theory of GMM and HMM, Maximum Expectedalgorithm, K-Means Clustering algorithm, Baum-Welch algorithm, Viterbi algorithmand Probability Maximizing, this paper completed the design of speech recognitionsystem based on GMM and HMM, completed the whole process of training andrecognition, Moreover, verified the performance and analyzed the two models under the different model parameters and different voice length by the experiments.(3) Based on the analysis of the SVM classification principle, researched andanalyzed the algorithm that extract the signal channel mode noise and its long timespeech signal characteristics on the speech, train and classify the features based onSVM, the experiment analyzed the feasibility of the algorithm.(4) Designed and realized the whole system of speaker recognition withanti-speaker audio replay attack detection. The system can complete the training andrecognition process based on the GMM and HMM models, achieve anti-recordingplayback attack detection based on the SVM, and achieve a combination of speakerrecognition and forward anti-replay attack and backward anti-replay attack. Andcomplete a performance evaluation that includes the feature extraction efficiency, thetraining efficiency, recognition efficiency and recognition rate.