Study On Blind Detection For Voices Disguise

In recent years,the wide application of digital audio editing software makes the voice disguise easier.At the same time,voice disguise presents an increasing tendency in illegal applications,which has a great negative impact on the estimation of the authenticity of voice for passive forensics.Thus,it is important to identify whether a suspected voice has been disguised or not.First,audio disguise and passive forensics are analyzed in this thesis,and then forged voices blind detection is studied to resolve the problem that the disguise of voices is limited in the non-noise environment.The main work is listed as below:(1)An algorithm which identifies the realness of voice is proposed based on LFCC and formant.The disguise voices essentially will modify the frequency spectrum of voices,which could be represented by LFCC and formant.LFCC and formant-based features are supposed to be effective for the identification of disguised voices.In this paper,LFCC statistical moments include mean values and unbiased variance,while the formant statistical moments only include mean value as acoustic features.Then,the extracted features and support vector machine classifiers are used to separate disguised voices from original voices.Under the environment of non-noise,babble,factory,volvo noises(SNR ? 15dB),the detection rate of the classifier is above 95%and the missing alarm rate is under 5%in only one database with different speakers,while under the environment of non-noise,babble,factory,volvo noises(SNR?15dB),the detection rate of the classifier is above 95%and the missing alarm rate is under 10%in different database.The experimental results show that the algorithm has a good performance on the blind detection of voice disguise in the actual noise environment.(2)Based on the feature that the kurtosis shapes of voices will not be effected by the introduced noises,a blind detection algorithm of voices pitch tamper is designed by the kurtosis statistical moments.Experimental results show that this algorithm has a missing alarm rate lower than 2%under different noise environment and have better detection rate under non-noise or noise environment with high SNR.