Audio Replay Spoof Attack Detection Using A GMM-RFPNN Model as Back-end Classifier

Abstract

Research on automatic speaker verification (ASV) techniques has received academic attention in recent years and has begun to be applied to authentication, but research on the security performance of ASV is just beginning. In this paper, we will focus on speech replay spoofing attack detection in speaker authentication techniques. Voice is a biological behavioral feature with high inter-class variability and susceptibility to environmental and temporal influences. In this paper, classical constant Q cepstral coefficient features (CQCC) and Gaussian super-vectors are used as front-end feature extractors and fuzzy polynomial neural network (FPNN) models with regularization processing are used as back-end classifiers for true and false speech detection. Compared with other traditional machine learning models and deep learning models, this model shows stronger robustness and generalization ability on acoustic environment and time variation, and good detection results can be obtained using a small number of samples for training. Tested on the ASV spoof 2017 version 2.0 database, the detection performance is improved by about 39% compared to the original baseline system.