A Study Of Speech Replay Attack Detection Techniques In Speaker Verification Systems

The speaker verification system is a one-to-one authentication system that has gradually started to be used in people's lives.This paper focuses on the detection of speech replay spoofing attacks in the security technology of speaker verification systems.In the research of speech replay attack detection,there are problems such as small scale,different contents and length of replayed speech data,as well as poor recognition of unfamiliar samples by the back-end detection model.In this paper,based on a detailed analysis of the changes in speech signals before and after replay,we propose for the first time to introduce fuzzy rules into the speech replay attack detection task and use fuzzy polynomial neural nets as a back-end detection model.In order to solve the problem of low recognition of unfamiliar samples by the back-end detection model,the use of support vector regression machine is proposed to replace the polynomial regression in fuzzy polynomial neural networks.The main innovations of the thesis are as follows.First,speech signals are analysed on sonograms,speech spectrograms and feature spectrograms,and it is experimentally demonstrated that the main focus of the replay attack detection task lies in the middle and high frequency bands of the speech signal.To address the problems of small speech sample data and different contents and lengths of speech signals in speech replay attack detection studies,the use of GMM-UBM model is proposed to extract Gaussian super vectors as the final back-end detection model input features.Second,to address the problem of high dimensionality of the extracted speech signal features,the algorithm of principal component analysis was used to perform dimensionality reduction operations.For the first time,fuzzy rules are applied to the field of speech replay attack detection,and fuzzy polynomial neural network is used as the back-end detection classifier,which improves the recognition ability of the model for unfamiliar samples to a certain extent.Compared with the baseline system,it possesses an improvement of 30.50%.Third,to address the problem of poor recognition of unfamiliar samples in speech replay attack detection tasks,a fuzzy neural network with a support vector regression machine as the neuron was designed based on a fuzzy polynomial neural network framework and a support vector regression machine.Compared with the baseline system,its recognition accuracy on the evaluation set possessed a 45.65%improvement.Experiments demonstrate that the use of a fuzzy polynomial neural network with a support vector regression fuzzy neural network can achieve good detection of unfamiliar samples in the replayed speech detection task compared to the officially provided baseline system.Experimental results on the evaluation set show that the detection model used in this paper has a higher ability to identify unfamiliar samples and performs well for different risk conditions.