基于带阻滤波倒谱系数的回放语音检测算法

A Replay Speech Detection Algorithm Based on Band-stop Filter Cepstral Coefficients

语音技术的发展及高保真设备的普及,使得回放语音给声纹认证系统的安全性带来了极大挑战;为此,本文提出一种基于带阻滤波倒谱系数的回放语音检测算法.该算法从特征层面和分类器层面对原始语音和回放语音进行了差异性分析,提出在时域使用带阻滤波的方法来保留高频和低频信号;然后将滤波后的信号进行倒谱特征提取并使用高斯混合模型进行分类;最后为验证本文提出的算法的有效性,我们将该算法在ASV spoof 2017数据库上进行了实验验证.结果 表明,我们提出的算法等错误概率为9.75%,相较于ASV spoof 2017数据库两种基线系统,本算法在等错误率方面分别有60%和20%的提升.

Voiceprint authentication system is seriously threatened by playback voice due to advanced voice technology and hi-fi equipment.Thus,it poses a great threat to the safety of the voiceprint authentication system.In this paper,we proposed a novel approach which can effectively detect the replay speech.First,we analyze the difference between the original speech and the replay speech in terms of feature and classifier.We employ the band-slop filter in the time domain to amplify the discriminative information.Afterwards,we calculate the cepstral coefficients of the filtering signal.Then the obtained cepstral features are fed into a Gaussian Mixed Model(GMM)for discrimination.Finally,the experimental results evaluated on ASV spoof 2017 database verify the effectiveness of our method,which can achieve an Equal Error Rate(EER)of 9.75%.Compared with two baseline systems of the ASV spoof 2017 database,our method has 60%and 20%improvement in EER,respectively.