多特征联合的语音被动取证方法

A multi-feature approach to passive-blind speech forensics

在分析篡改音频特征变化的基础上,提出了一种语音被动取证方法。采用语音的美尔倒谱域参数及其动态特征参数和小波域统计矩特征来建立模型,并选取支持向量机(SVM)作为分类器以寻找最优分类平面,实现对可疑语音信号真实性的盲取证。实验结果表明,该方法对语音片段的删除、剪接和替换等改变语音内容真实性的篡改操作能够达到较高的检测准确率。

A blind detection method for speech forensics is proposed based on hybrid domain spectral analysis. As forgeries lead to changes in dynamic property and high-order statistics, joint features in both mel-cepstrum and wavelet domain are adopted for classification. A SVM-based classifier is employed to find the optimal threshold for forgery detection. The validity of this scheme is evaluated under forgeries like deletion, splicing and substitution. Simulation results show that this method achieves promising accuracy for passive-blind speech forensics.