基于中心对称局部二值模式的合成伪装语音检测方法

针对基于局部二值模式的伪装语音检测方法的合成语音检测准确度较低的情况,提出了一种基于中心对称局部二值模式的伪装语音检测方法。该方法通过短时傅里叶变换得到语音信号的语谱图,再利用中心对称局部二值模式提取语谱图的纹理特征,并用该纹理特征训练随机森林分类器,从而实现真伪语音的判别。该方法综合考虑语谱图中像素点的数值大小和位置关系,包含了更加全面的纹理信息,并将特征维度降低至16维,有利于减少计算量。实验结果表明,在ASVspoof 2019数据集上,与传统的基于局部二值模式的伪装语音检测方法相比,所提方法将合成伪装语音的串联检测代价函数(t-DCF)降低了16.98%,检测速度提高了89.73%。

Synthetic spoofing speech detection method based on center-symmetric local binary pattern

In view of the fact that the local binary pattern (LBP) based speech spoofing detection method has low detection accuracy when detecting synthetic speech, a spoofing speech detection method based on center-symmetric local binary pattern (CSLBP) was proposed.In this method, the spectrogram of the speech signal was obtained through short-time Fourier transform (STFT), and then the texture feature was extracted from the spectrogram using the CSLBP.The random forest classifier was trained by the extracted texture feature to realize the discrimination of genuine and spoofing speech.The CSLBP-based method comprehensively considered the value and position relationship of pixels in the spectrogram so as to contain more texture information, and reduced the feature dimension to 16 beneficial to decrease the amount of computation.Experimental results on the ASVspoof 2019 dataset show that, compared with the LBP-based spoofing detection method, the proposed method reduced the tandem detection cost function (t-DCF) of synthetic spoofing speech by 16.98% and increased the detection speed by 89.73%.