关键词: 卷积神经网络, 翻录语音检测, 语谱图, 录音设备, 网络框架

Abstract: With the popularity of portable and high-fidelity recording devices, the high similarity between the recaptured voice and the original voice may be utilized for attack the speaker verification system illegally by the criminals, which harms the legitimate interests of the system user.In order to improve the ability of resisting this attack, this paper proposes a CNN-based algorithm of detecting recaptured speeches by using spectrogram of the voice. For the detection task, a novel network structure is constructed, and the effect of the spectrograms with different window shifts are discussed.In addition, the cross-over experiments for various eavesdropping and replay devices are construct in this paper.Experimental results demonstrate that the proposed method can accurately discriminate whether the recording voice is recaptured or not. The detection rate achieves 99.26%, and furthermore is higher than of the state-of-the-art methods.

Key words: convolutional neural network(CNN), recapture voice detection, spectrogram, recording devices, network structure