基于对角双谱系数的录音回放攻击检测

Replay Spoofing Attack Detection based on Diagonal BiSpectrum Coefficients

录音回放攻击是说话人确认系统实际应用时面临的最大挑战。为实现录音回放攻击检测,首先对录音回放语音的输入信道结构特点和频率响应进行分析,发现输入信道结构上的差异会导致时域卷积效应和频率响应的高频衰减。因此,基于双谱分析的高维度频域信息表达能力,结合对角切片双谱计算和IMel滤波,提出一种录音回放攻击检测的对角双谱系数(Diagonal BiSpectrum Coefficient,DBSC)特征。在基于ASVspoof 20172.0语料库的录音回放攻击检测实验中,采用DBSC特征的录音回放攻击检测等错误率(Equal Error Rate,EER)为22.21%,相比常数Q倒谱系数(Constant Q Cepstral Coefficient,CQCC)特征和Teager能量倒谱系数(Teager Energy Cepstral Coefficient,TECC)特征分别下降了6%和38.39%。结果说明,DBSC特征能够有效表征真实语音和录音回放语音的差异,具有良好的录音回放攻击检测能力。

Spoofing attack is the biggest challenge in the practical application of speaker verification system.In order to detect spoofing attack detection,the input channel structure characteristics and the frequency response of spoofing speech is first analyzed,and it is found that the input channel structural differences can lead to time domain convolution effect and high frequency attenuation of frequency response.Therefore,based on Diagonal BiSpectrum analysis,high dimensional frequency domain information expression ability difference between genuine speech and spoofing speech is more visible.Combined with the Diagonal Slice BiSpectrum and IMel filtering calculation,Diagonal BiSpectrum Coefficient(DBSC)is put forward to detect spoofing attack.In the spoofing attack detection experiment based on the 2017 ASVspoof Challenge Database 2.0,the Equal Error Rate(EER)of DBSC is 22.21%for the spoofing attack detection system.Compared with the baseline system using Constant Q Cepstral Coefficient(CQCC)and Teager Energy Cepstral Coefficient(TECC),the relative decrease is 6%and 38.39%respectively.The results indicate that DBSC feature can effectively reflect the difference between genuine speech and spoofing speech,and has good spoofing attack detection capabilities.