Deep Learning as a Speaker Spoofing Countermeasure

作     者：Heinrich Dinkel 

作者单位：上海交通大学 

学位级别：硕士

导师姓名：Kai Yu;Yanmin Qian

授予年度：2016年

学科分类：0711[理学-系统科学] 12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 07[理学] 081104[工学-模式识别与智能系统] 08[工学] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

摘      要：Recent development of algorithms and hardware for speech applications has enabled researchers to introduce deep neural networks（DNN）as a state-of-the-art machine learning framework,drifting apart from traditional Hidden Markov Model（HMM）and Gaussian Mixture Models based *** to the traditional models,DNNs can be deployed in nearly any *** the huge success in image and speech recognition tasks,the full potential within the speaker identification community has not been explored *** the speaker identification community focuses on building potent speaker discriminative systems,a new threat to speaker identification was recently described: attacks in form of spoofed *** utterances can severely harm a preexisting speaker identification system to the point at which unauthorized users can trespass into the *** research,which focuses on combating this thread concluded into two spoofing challenges: ASVSpoof2015 and *** challenges are the first to provide sufficient data to investigate methods,which can prevent these spoofing *** thesis focuses on the construction of deep neural network based spoofing *** neural networks have already been used in this field,they mostly act as feature extractions,marginalizing their classification *** this work,we revise the deep feature extraction framework and propose two approaches to the spoofing problem: An End-to-end system,capable of directly accepting or rejecting an utterance from raw waveforms and a small scale convolutional network,outperforming any preexisting neural network approach by large,while using significantly less parameters and training *** proposed end-to-end CLDNN model outperforms the best current result on the BTAS2016 corpus of 1.21 % half total error rate（HTER）to 0.82 % *** addition the small-footprint DCNN model boosts the previously best neural network based model’s result of 1.1 % down to 0.7 % equal error rate