A Noise-robust Algorithm For Spoofing Speech Detection Based On AMBP And RAMBP

Spoofing speech detection is to analyze and process the speech to determine whether the speech belongs to a real speaker or a spoofing speech.Nowadays,the information processing technology is developed,the spoofing speech is easy to conduct,the vulnerability of Automatic Speaker Verification(ASV)in the face of spoofing speech is gradually exposed.Spoofing speech is a speech that is similar to the target speaker in all aspects generated by some technical means.There are four types of spoofing speech: Impersonation,Replay,Voice conversion and Speech synthesis.These spoofing speeches will do great harm to the security of ASV.The purpose of the spoofing detection system is aim at detecting and rejecting spoofing speech to improve the security of the ASV.The texture feature of speech signal is an important and effective feature parameter to distinguish real speaker’s speech from spoofing speech.The Local Binary Pattern(LBP)based spoofing speech detection system extract the texture feature vector from the acoustic features which is extracted from the speech signal in advance for spoofing speech detection.The detection performance of this method is related to the acoustic features extracted.If the spoofing speech changes the acoustic characteristics,the detection performance of the system will be poor.Moreover,the performance of spoofing speech detection becomes worse when the system is under noise environment.To address the issue that the noise robustness of spoofing speech detection algorithm based on LBP is poor,a noise robust spoofing detection method based on Adaptive Median Binary Pattern(AMBP)is proposed.The variable Q transformation is used to convert speeches into time-frequency representation.AMBP was used to extract texture feature vectors from the spectrogram of the speeches in training dataset for training classifier,so as to conduct the detection of spoofing speech.The experimental results show that the AMBP based spoofing speech detection system can achieve better detection performance in a noisy environment,and the higher the SNR,the better the performance.However,under the condition of low SNR,the performance of this algorithm still needs to be improved.In order to enhance the noise robustness of the whole system,especially the system performance under the condition of low SNR,a spoofing speech detection algorithm based on Robust Adaptive Median Binary Pattern(RAMBP)is proposed.The object of feature extraction by the RAMBP algorithm is also the spectrogram.Based on AMBP,the noise detection module was introduced to mark the polluted pixels in the spectrogram and eliminate their influence in the process of feature extraction.The texture feature vector is obtained after threshold generation and binary patterns generation,which is used for spoofing speech detection.The experimental results show that the RAMBP-based spoofing speech detection system has better detection performance on the whole system.Especially under the condition of low SNR,the RAMBP detection system has more obvious advantages in noise robustness than the AMBP-based spoofing speech detection system.