采用完整局部二进制模式的伪装语音检测

Completed local binary pattern based speech anti-spoofing

针对基于局部二进制模式(LBP)的伪装语音检测算法在检测语音转换的欺骗攻击时效果较差的情况,提出了一种基于完整局部二进制模式(CLBP)的伪装语音检测方法。利用变量Q变换(VQT)生成语音信号的语谱图并应用完整局部二进制模式提取语谱图中的纹理特征向量,再用该纹理特征向量训练真/伪语音分类器,实现伪装语音检测。实验结果表明,所提方法在检测语音转换的欺骗攻击时效果更好,并且当变量Q变换的附加参数γ值为50时,由完整局部二进制模式特征向量训练的基于径向基核函数的支持向量机具有最佳的伪装语音检测性能。

In view of the fact that the disguised speech detection algorithm based on local binary pattern(LBP)is not effective in detecting the spoofing attack from voice conversion,an anti-spoofing method based on completed local binary pattern(CLBP)was proposed.In this method,the spectrogram of speech signals is generated by the variable Q transformation(VQT)and used to train the true/spoofed speech classifier,so as to perform the detection of disguised speech.The experimental results demonstrate that the proposed anti-spoofing method based on the CLBP in the detection of voice conversion deception is better than the LBP-based algorithm,and when the parameterγin VQT is set to 50,the detection system based on CLBP and SVM-RBF has the best performance for anti-spoofing the disguise speech.