Study On Content-based Blind Detection For Disguised Voice

With the rapid development of Internet technology,communication through Internet has become the mainstream form.For its simplicity of using and acquisition,audio is regarded as one of the main communication medias.However,the emergence of audio editing software poses the serious threat to the security application of audio file.The audio content could be disguised by audio editing software easily,such as insertion,deletion,pitch-scaling,transform voice of speaker,etc.Disguised audio becomes more and more difficult to be distinguished through human ears.To solve these problems,blind detection of disguised audio technology has been researched and improved by many scholars.Different from the audio active forensics,blind detection of disguised audio which known as audio passive forensics relies on the information in the receiving end completely,and detects disguised audio without pre-embedding information.The authenticity and integrity of audio could be effectively protected.Voice signal has a wide range of applications as the audio signals,such as military voice commands,VoIP orders,voice evidence of court,etc.Blind detection of disguised voice technology also has a high practical application value and has been gradually received the attention of many scholars in recent years.In this thesis two blind detection algorithms are proposed,aiming at disguised voice by audio editing software.The algorithms are designed for pitch-scaling voice and voice-changer voice.The main work is as follows:1)An algorithm of cochleagram-based identification with pitch-scaling is proposed.The algorithm combines speech enhancement technology with cochleagram to achieve the detection of pitch-scaling.The proposed algorithm is briefly introduced as follows: First,the voice signal is passed through the Least Mean Square(LMS)filter.Then the cochleagram is extracted and calculated multi-resolution to construct the Least Mean Square-Multi Resolution Cochleagram(LMS-MRCG)for pretreatment.The Universal Background Model(UBM)is used as the classifier for detection.The experimental results show that the algorithm can detect voice with pitch-scaling effectively and classify types of pitch-scaling accurately.In the non-noise environment,the detection rates up to 97.50%.And the algorithm has good anti-noise performance.In the noise-added environment,especially in low SNR environment,the detection rate can still be maintained above 85.83%.2)An algorithm of identification of voice disguised by voice changer is proposed.Compared with pitch-scaling,voice change has more variable factors.In the proposed algorithm,first the cepstrum-based endpoint detection is used to distinguish the segments of voice and no voice.Then features are extracted from voice segments.In feature extraction,the algorithm selects the fundamental tone rates combined with Gammatone Feature(GF)generated by cochlear simulation model,first-order and second-order difference value of GF to construct combination feature.Gaussian Mixture Model(GMM)is selected as classifier for detection.The experimental results show that the overall detection rate of voice disguised by voice changer can reach 98.33%.Detection rates for each type of voice change could be maintained above 87.8% in the noise-added environment.Therefore,the proposed algorithm is an effective algorithm of detecting voice change.