Authentication And Recapture Detection Of Recordings

With the popularization of speech signal digitization,it is very convenient to store,transmit and share the speech signal data.At the same time,simple and powerful audio editing software has developed rapidly.It is an easy job to record a voice,or edit to polish speech signal.While these technologies bring people conveniences,they may cause some security problems.For example,using audio processing technology can change the content of a speech or generate a forged speech recording.Once some people use this forged speech recording for illegal purposes,it will cause a threat to the safety of people's lives and properties.Therefore,it is of great significant to detect the authenticity of digital speech signal.Although there are a lot of digital audio forensics researches,it still can't meet the requirements of the society and people.In this thesis,the authenticity of the recording and the detection of the recaptured digital speech signal are studied,and the main contents are as follows:1)The authenticity of the recording.With the popularity of audio editing software,people can use some common functions in the audio editing software(such as filtering and reverberation)to beautify and modify the audio.These simple and powerful audio editing software bring convenience to people,but also provide a chance for some lawbreakers.For example,some people could use filtering operations to achieve a smooth effect on the spliced audio,which could conceal the splicing traces.If this kind of audio appears at the court trial and is accepted,there is no doubt that it would have an impact on the justice.In addition,some people could simulate other people's voice by the modulating function to achieve phone fraud.This will result in serious threaten to people's assets.As you see,it is very necessary to detect the authenticity of the speech signal.Inspired by the co-occurrence matrix of the digital image for reference,this thesis presents the amplitude co-occurrence vector(ACV)feature which can apply to the speech signal.For the feature extraction,the speech signal will be quantized,then the probability distribution of the co-occurrence vectors,which is formed by several adjacent sample points,will be calculated.This feature reflects the fluctuation characteristics of the adjacent sample points,which has a good effect on the detection of the processing speech signals,and the accuracy of the experiment can reach 95%.In addition,we conduct experiment to distinguish twelve kinds of processing operations from two audio editing software and the results show that the feature can be used to identify the types of processing functions.2)The detection of the recaptured speech signals.The recaptured operation not only can forge the scene in the speech signal,but also can be used to attack the identity authentication system,which is based on voice characteristics.So it is also important to detect the recaptured operation.In this thesis,we modified co-occurrence matrix features by the statistical analysis point of view,and use extended ACV features(X-ACV)to detect recaptured speech signal.In order to make it more suitable for recaptured detection,we analyze the quantized threshold value and introduce sample combinations of different sample interval.We also construct a database of recaptured speech signals,which includes different recording equipments and scenes.Such database provides sufficient data for the experimental part.In Comparison with the MEL-cepstrum coefficient feature and ACV feature,experiments show that X-ACV feature has better detection performance in the detection of recaptured speech signals and the accuracy rate can reach 96%.Then we divide the database into different scene to form sub-dataset.The accuracy rate for the detection in the same scene can reach 99.36% and for the detection in the different scene can reach 95.69%.