Proceedings oI International ConIerence on Computing Sciences WILKES100 ICCS 2013 ISBN: 978-93-5107-172-3 A passive approach to detect copy-move Iorgery in digital speech audio signal Abhishek Verma 1,* , Vivek Kumar Singh 2 and R.C.Tripathi 3
1 M.Tech., Speech and Image Processing Lab, Indian Institute of Information Technology Allahabad 211012, Uttar Pradesh, India. 2 Research Scholar, Speech and Image Processing Lab, Indian Institute of Information Technology Allahabad 211012, Uttar Pradesh, India. 3 Professor,Department of Human-Computer Interaction, Indian Institute of Information Technology Allahabad 211012, Uttar Pradesh, India Abstract Due to the advancement in the digital technologies, digital tampering becomes very easy in the multimedia contents like digital image, digital audio signal etc. Recent news related to Iorgery in speech clips have already attracted the attention oI mass in India as well as in the other countries. Using advanced and powerIul audio editing soItware and tools, the digital audios tampering becomes very common now-a-days. In digital speech tampering, some words or part oI sentence are deleted or inserted in the same signal or in diIIerent signal and sometime can be mixed with other signal oI the same person at diIIerent time. In this paper, the Iocus is to detect 'copy-pasted also called 'copy-move type oI Iorgery. In this type oI Iorgery, one portion oI the signal (either a word or a part oI sentence) is copied and then pasted somewhere else in the same digital audio signal. The aim oI this Iorgery is to alter the useIul inIormation in the original signal. Some most common type oI Iorgery cases are those in which the audio clips are the main case evidence, which requires that the case evidence should be authentic. In this paper, we proposed a method to detect the copy-move Iorgery in digital speech clips. By using Ieatures oI the audio signals, we examine the whole audio signal. A window-by-window basis approach is used Ior this. In result oI this paper, we show the result with tampered speech clips by highlighting those part oI the speech signal clips which are suspected to be Iorged. 2013 Elsevier Science. All rights reserved. Keywords: Digital speech Tampering, Copy-Move Forgery, Audio Features, Neighbor Shift matching 1. Introduction Digital audios have become ubiquitous with the popularity oI the internet and portable digital devices such as personal music players and smart phones. In the meanwhile, rapid developments oI low-cost and sophisticate editing soItware make the modiIication oI audio Iile much easier Ior untrained users. Digital Iorgery is a kind oI threat in which the digital content is modiIied, altered or tampered in a way to disorient the people Ior any beneIit involved. With the advancement in the internet technologies it is now very easy to collect and store, as well as upload the digital multimedia contents to or Irom the websites. The easy and Iree availability oI audio editing soItware, the alteration in any manner in the audio becomes very easy. In this way, the digital audio Iorgery becomes a serious problem and to detect this problem is itselI a challenge in the present scenario. There are various cases about the Iorgery in the digital content spreads across all over the world. 184 Elsevier Publications, 2013 Abhishek Verma, Vivek Kumar Singh and R.C. Tripathi Figure. 1 Following are the some high proIile cases which was/are in controversy: 1. Amar Singh controversial telephonic record releases in the year 2011. (Not prooIed) 2. American Actor Mel Gibson`s controversial audio recording in the year 2010. 3. Bin laden`s tape not genuine` the tape released in 2002 was told that the voice in the tape is not the voice oI bin laden. A huge number oI Iorged audios calls Ior more eIIective tools Ior the authentication and Iorgery detection Ior digital audio 1.1 Types of forgery in the audio contents Some common types oI Iorgeries in the digital audio/speech signal are:- Insertion: Here, some part or a complete diIIerent audio clip is inserted at any place in the original audio clip. Copy-Move: Copy-move Iorgery is a type oI Iorgery in which some part oI the signal is copied and then moved/pasted to diIIerent location in the same signal. This type oI Iorgery is considered to be detected in this paper|Iig 1| Deletion:Some part is deleted Irom the audio/speech signal/clip. Substitution:The part oI some other signal is substituted in the original signal at the place oI a particular word while the original signal size does not modiIy. Splicing:A kind oI mixing is done between two audio/speech clips. Sections Irom one audio are inserted to another audio 1.2 Previous work Forgery becomes very common and serious type oI threat. A noticeable amount oI image Iorgry detection techniques are already been in use. Area oI detection oI attacks on audio is in its starting phase. Math et al. |1| shows a hierarchal overview oI various kinds oI digital Iorgeries. Lineage (Ilow) mapping and source (origin) identiIication approaches are mentioned to detect the Iorgery in digital content. While Greige et al. |2| discussed biometric identity veriIication problems in audio-visual biometric systems which select suitable audio signal Ieatures and extract Irom the questioned speech signal. A determined number oI Mel Irequency cepstral coeIIicients (MFCC) are used to represent the spectral envelope oI the speech signal to veriIy the speaker. Block matching is a very popular method to detect the copy-move Iorgery in the digital image. it is used in the method proposed in |3,5,10|.Singh et al. in |3| proposes a eIIicient ad robust method to detect copy move type oI Iorgery which is the basis to our approach. In the same, Ieatures are calculated Ior the diagonally divided blocks oI the image and then using lexicographical sorting and neighbor shiIt matching the suspected region which are copied and pasted are marked, in this way ,it detect region duplication in the image. Lin et al. |5| represent an integrated technique Ior splicing and copy-move Iorgery detection in image . Images Irom RGB color space are converted into YCbCr color space. Splicing is detected by computing the probability oI the Iorged block using statistics oI the histogram which again is computed by checking the DQ eIIect oI this histogram. For detection oI copy-move Iorgery, SURF algorithm is used. Bayram et al. |10| uses block-by-block approach as in 185 Elsevier Publications, 2013 sm,i. .i c.,,x.. l...., A assive pproach to etect opyMve orgery in igital peech udio ignal |3|. The duplicated regions which are candidate oI copy-move Iorgery are detected by calculating Ieature vector oI the image block-by-block basis. Lexicographical sorting and bloom Iilters with suitable distance measures are then applied to ensure about the copy-move Iorgery in the image. Except this the Iorgery work in the audio signal are very less and still in progress and one oI the research area in the digital signal processing Iield. However, the work in the Iield oI digital audio Iorgery detection done by Yang et al.|8| whose proposed method is to detect tampering in MP3 audio Iormat by using number oI active (non-zero) spectral coeIIicients as Iunction oI Irame oIIsets. Korycki et al.|9| does the detection oI tampering by doing visual inspection oI the audio signal in time-Irequency domain, the possibility oI tampering is decided. The short-time Fourier transIorm and the analysis based on ENF Irequency criteria is used. Maher et al.|4| explain ENF(electric network Irequency) criteria, Time-domain level detection, Irequency- domain Iiltration like common digital signal processing approaches .Other manual process oI tampering detection like Physical examination and inspection oI audio recordings ,critical listening and waveIorm observation, and ,careIul listening till the end oI the recordings are also mentioned which are also used Ior analog recording oI the audios. Koenig et al. |11| shows various approaches Ior Iorensic examination oI the audio. Spectrographic Analysis SoItware, Fast Fourier TransIorm (FFT) Analyzers and SoItware, High-Resolution WaveIorm Analysis SoItware etc.are described and explained. Iorensic evidence examination and procedures are given like Evidence marking, Physical inspection, Digital data imaging, Playback/conversion optimization, Critical listening, High- resolution waveIorm analysis, Narrow-band spectrum analysis, Spectrographic analysis, Digital data analysis, Miscellaneous techniques, Work notes and reporting. Digital Audio signal Ieature plays very important role in speaker veriIication, audio matching and authenticity oI the audio signal. In |6, 7| various audio Ieatures are discussed. Mitrovic et al. |6| proposes various audio Ieatures which are useIul in audio classiIication and matching process. The Temporal Ieatures, Physical Irequency Ieatures, Perceptual Irequency Ieatures, Cepstral Features, Modulation Irequency Ieatures, Eigen domain Ieatures, Phase space Ieatures are discussed which have numerous applicability in the audio and speech signal processing. While Herre et al. |7| those audio Ieatures are proposed which are useIul in audio matching. The Ieatures are described are spectral peaks, spectral Ilatness measure, spectral crest Iactor. These mentioned Ieatures are well known Ior their predictability and robustness which doing matching oI the two given audios. We use these Ieature in our method also. Except this |12,13| does a small survey in the audio Iorensic Iields. Gupta et al.|12| discussed recent developments in Iorgery detection oI digital audio Iiles by elaborating the audio processing systems with Iorgery detection. Techniques like spectrogram analysis, device based authentication, environment based authentication are discussed. Kraetzer et al. |13| presents approaches Ior digital media Iorensics to determine the used microphones and the environments oI recorded digital audio samples by using known audio steganalysis Ieatures .AAST and WEKA data mining soItware are used Ior Ieature computation and classiIication. In this paper, we describe a method that can be applied to detect a common Iorm oI tampering in digital audio signals known as copy-move Iorgery, where sections in one audio is copied and pasted into other location in the same audio. Our method achieves this by detecting matching pairs oI windows over the signal by comparing audio Ieatures using shiIt vector calculation. 2. Features Used In digital audio signal processing, there exists class oI parameters (Ieatures) like low-level signal parameters, MFCC, psychoacoustic Ieatures including roughness, loudness and sharpness etc.There exists a class oI statisticalIeatures oI audio signal which are known Ior their robustness and low computation overload. These statistical Ieatures perIorm well in matching process. Some are as Iollows which is useIul in the proposed method Root mean square value (RMS value): -It is given as:- Where, x(n) is the signal`s sample value at n,N is the length oI the signal and is the Root mean Square value.It is useIul Ior sinosoids, means where the variates are positive and negative. It will give the value equal to the one period oI the signal. 186 Elsevier Publications, 2013 Abhishek Verma, Vivek Kumar Singh and R.C. Tripathi Zero-crossing rate (ZCR): The rate at which the sinusoids` sign change occurs is called zero-crossing rate. It is deIined as:- Where, is the signal sample value at i, sign gives sign value at ith position, Is is the sampling rate oI the signal and N is the signal length taken or given, This Ieature has been used heavily in both speech recognition and music inIormation retrieval, being a key Ieature to classiIy percussive sounds Spectral Flatness measure (SFM) : Spectral Ilatness is calculated by dividing the geometric mean oI the power spectrum by the arithmetic mean oI the power spectrum.It is Iormulated as:- Where, denotes the power spectral density (PSD) coeIIicients in Irequency band, N is the signal length. SFM shows robustness in the matching process between the audio signal pairs. Spectral Crest Factor (SCF): It is the value obtained by taking the ratio oI the largest power spectral density (PSD) and the mean power spectral density (PSD) value in the Irequency band. Where, denotes the power spectral density coeIIicients in Irequency band, N is the length oI the signal. These two spectral Ieatures SFM and SCF have the property that they are not signiIicantly inIluenced by the speech coding system or by other intended modiIications oI the audio signal Mel-Frequency Cepstral Coefficients (MFCC): Mel-Frequency Cepstral is the representation oI the short-term power spectrum oI the sound, and Mel- Irequency Cepstral CoeIIicients are those coeIIicients which make the MFC oI a sound. This Ieature is very useIul and popular in speaker recognition and veriIication systems. While in our system we are Iocusing on the matching pairs which have same vector oI Ieature values. So, in this way MFCC is used only Ior matching purpose not Ior the recognition here. Power Spectral Density (Spectra Density Mean and Spectral Density Standard Deviation): Power Spectral Density (PSD) describes how the power oI a signal or time series is distributed with Irequency. Power Spectral Density (PSD) coeIIicients oI the signal are calculated using the Iormula:- Where, denotes the power spectral density (PSD) coeIIicients in Irequency band,Nw denotes overlapped segments oI the x(n)where x(n) is divided into Nd length window w(n) where w(n) is the window chosen (hamming window) and the Iormula inside , , is nothing but the Discrete Fourier transIorm (DFT)oI the signal. 187 Elsevier Publications, 2013 A assive pproach to etect opy-ove orgery in igital peech udio ignal SFM and SCF are computed Irom the power spectral density. In this way the statistical data Irom the PSD like mean, standard deviation, min, max oI the PSD can be helpIul. So considering this Iact, Spectral Density mean and Spectral Density Standard deviation are used as Ieatures Ior the proposed method 3. Proposed Method Figure 2. Flow Chart oI Proposed Methodology. The proposed method takes the input digital audio speech signal and then it starts processing in the Time- domain. To reduce the size oI the input signal, down sampling can be done. AIter this, the signal is divided into overlapped windows oI same size .Then these windows are matched to detect iI there is any copy-move Iorgery present in the signal or not. Detailed procedure is divided into Iollowing parts:- DownSampling (Optional):-While representing the signal into its Time-domain, it has a large number oI sampling values oI the signal, which are to be processed and take much time to process. So, Downsampling is done with Iactor oI 1 on the signal to reduce the signal size assuring the nyquist rate Iormula. Because, there exists the problem oI losing those sampled values in the signal which may be tampered, not downsampling can be ignored Windowing of samples:-AIter step 1, the windowing process starts .Consider M samples in a window with N overlapped sample values between two adjacent windows oI the signal; here every window is treated as a signal itselI. So Ior a signal oI length L, there will be WL/M, iI N0; 188 Elsevier Publications, 2013 Abhishek Verma, Vivek Kumar Singh and R.C. Tripathi WL-(M-(M-N)), iI NM-1; Windows in the given signal. For example, II our signal length is 30,000,and the value taken Ior M is 1000 samples and N999 overlapped samples between the adjacent windows, then total number oI windows Ior the given signal will be, W30,000-(1000-(1000-999)) W29001 Feature Extraction/computation: - The next step is to extract/compute Ieatures discussed in section 3 Ior every window by considering every window itselI as an individual signal. Let Bi denotes the ith window which will contain the Ieatures Ior the ith window, Total there are Bi (i1, 2....W) where W is the total number oI windows Ior the given input signal. Also with the Ieature set Fij (i1, 2...W, j1, 2..7), a vector oI the Iorm, (Fij, Bi) Where Bi is the window number and Fij is the vector oI seven parameters values; is generated. Features are extracted/calculated directly Irom each window, the computation as Iollows. From the experiment it is shown that Root mean Square (RMS) and Zero-Crossing Rate (ZCR) have the exactly similar values Ior the matched pair, For Fij(j1), Iirst Ieature which is computed/extracted is RMS and then For Fij(j2) ZCR is computed. AIter the calculation oI the two statistical Ieatures, For Fij (j3), computation Ior MFCC starts. First coeIIicients oI the resulting MFCC array Ior the window is taken. Because, MFCC shows the uniqueness to the speech, it will give almost similar coeIIicients values Ior the same speech signal, so it is computed as the third Ieature. Now, two Spectral Ieatures SFM and SCF are computed as Fij (j4 and j5), as already mentioned that these two are known Ior their robustness and these two are also used in the audio matching process. For SFM and SCF, power spectral density (PSD) coeIIicients are used.AIter this the Ieature computation process ends with computation oI the last two Ieatures. calculate mean and standard deviation oI the PSD as the Fij (j6 and j7) Ieature.Thease two Ieatures values have less ability to give the inIormation about the matched pairs comparatively, but iI thrsholding oI very narrow range values is used then chances oI getting the almost similar Ieature value are increases .Now, store set (Fij, Bi) in a multidimensional array, A, where array A has the rows equal to the number oI the windows and have eight columns. A contains all the real values. The Array A has the Iorm like:- Fig 3. Structure oI the array AIter calculating all the Ieatures Ior a window, matching process starts. Lexicographical Sorting:-Now, all these are stored in a multidimensional array A. Every row oI this array has eight values; one window number which is starting sample value oI the window and other seven are Ieature values described above. This means that ,this array has W number oI rows and eight columns First seven 189 Elsevier Publications, 2013 A assive pproach to etect opy-ove orgery in igital peech udio ignal column are used in sorting ,highest priority is given to the Ieatures I4 and I5 then I1 and I2 and then all other Ieatures gets the priority. Detection of suspected copied region:- Array A is successIully sorted, hence the rows, representing windows with very similar Ieatures, are arranged nearby. Those matched pairs or set oI pairs which are the candidate oI possible Iorged part are taken out by add/subtract preset threshold value which are diIIerent Ior diIIerent Ieatures set, they are, +0.001 threshold Ior the MFCC, SFM and SCF Ieatures value, +0.007 Ior standard deviation oI PSD values and +0.01 Ior mean oI PSD values. In all the experiments, the given threshold is used and except some results almost oI the tampered signal detected. Neighbor Shift matching: - It is very much possible that the nearby windows detected aIter step 5 are those matching pair oI the windows which are just adjacent to each other. Neglect and discard these matching pairs because this part cannot be considered as the Iorged part, means, the windows are just 1 sample overlapped and one sample Iorgery is meaningless and just one sample cannot show a word or alphabet in the signal. In Neighbor shiIt match, those (Fij,Bi) vector which shows a long chain oI corresponding matching windows are considered and other vectors discard Irom the array we get in step 5. Neighbor ShiIt value is calculated by subtracting two corresponding Ieature vectors. ShiIt vector oI the entire suspected duplicate region will be same. |3| As in 1. ShiIt ( (j), (j)) ShiIt ( (k), (k)).1 Two copied and then moved areas will yield almost same pair oI identical Ieatures. This Ieature will yield same shiIt vector. This shiIt vector will be checked Ior a particular number oI neighbors. Same shiIt vector will be showing the duplicated region. The Iollowing result shows the eIIect oI the neighbor shiIt matching Fig 5. BeIore neighbor shiIt matching Fig 6. AIter neighbor shiIt matching Detection and marking of Duplicated Region:-Now, those windows which are suspected to be tampered are in hand. AIter step 6, the decision about tampered region can be taken. The proposed method mark tampered region in the time domain plot oI the signal in the result. The two marked regions shown in result tells that one oI them is copied and pasted at the other place but can not tell which one is copied and then pasted. Except Ior some Ialse matching, the system detects the tampered region successIully. The overall procedure is as Iollows in a Iorm oI algorithm; 190 Elsevier Publications, 2013 Abhishek Verma, Vivek Kumar Singh and R.C. Tripathi 3.1 Algorithm-copy-move forgery detection of digital audio speech signal:- Let x(n) be the digital speech audio signal represented in Time-Domain. M is the length oI window set by user; L is the length oI the signal. Fij ((i1, 2, 3...W) (j1, 2.7)) is the Ieature vector Ior every window. Bi is the ith window and i is the Iirst sample location as well as the window number.W is the total number oI windows Ior the input signal computed by the Iormula given in section 4.N is the number oI samples which are overlapped between the adjacent windows, A is the multidimensional array oI size (W 8), where ith row oI A is a vector (Fij, Bi) oI length 8, Now the algorithm Ior the proposed method is as Iollows:- //Set the value Ior M and N Initialize M, N //Determine total number oI Window we will have in the given signal Calculate W //Create and Initialize multidimensional array Create and Initialize Array A //Do windowing and calculate/extract Ieature one by one Ior every window Table 1. Copy-Move Iorgery Detection Algorithm Begin For i1 to L Get window Bi For j1 to 7 Extract/calculate Ieature set Fij Ior Bi Store (Fij, Bi) in A End End For i1 to W Do Lexicographical sorting oI A End For i1 to W Thresholding oI the window Ieatures Ior those suspected pairs End For i1 to W Do Neighbor shiIt matching process over the Array A End Decide and Mark End 4. Experimental result/Description The proposed Copy-move Iorgery detection algorithm presented in Section 5 and section 6 was implemented with MATLAB 7.10.0.The implementation is done on a computer oI CPU 3.0GHz with main memory 4GB and secondary storage memory oI 460 GB. We collect a set oI 100 high qualities audio speech signals in the with 32 KHz sampling rate oI length around 4-5 seconds Irom the TIMIT dataset |15|.The speech clips are recorded in the voice oI either male or Iemale but not both .The Forged signals are made by the popular audio editing soItware (like audacity, cool edit etc.).We made 100 copy-moved Iorged signals corresponding to the 100 original signals we have chosen Irom the dataset. Except the copy-move Iorgery, all the test audio clips are Iree Irom any other type oI audio tampering discussed in section 1. For all the results, we have taken 1000 sample size Ior the window with 999 samples overlapped to the two adjacent windows. Downsampling is also done Ior some audio clips.Thresholding is used whenever needed to 191 Elsevier Publications, 2013 A assive pproach to etect opy-ove orgery in igital peech udio ignal retrieve those matched window chain pairs which becomes very similar to each other, however the threshold value range taken is very small. For example, this audio clip is recorded in a Iemale voice. The sentence oI the clip is 'guess the question from the answers in original TIMIT |15| data set. We Iorged this audio clip as 'guess the question from the answers guess and checked Ior the tamper detection which is surely present. Without doing downsampling oI the Iorged clip we start the procedure. The window size is taken 1000 sample and overlapped size is 999 sampled between the adjacent windows. There was no other manipulation done /present in the clip going to experimented. The procedure takes exceeded time but we give the priority to the result`s accuracy, no matter how much time it takes. Various attempts were made with diIIerent threshold and diIIerent shiIt vector counts. The process repeats more than one time on the audio clip to gain accuracy. Besides some Ialse detection oI the result, we were able to decide about the Iorgery attempt done on the given audio by marking those regions which were suspected to be tampered Fig 6a. Iorged audio clip Fig 6b. Detection Results The two rectangular boxes shows us the suspected part which were copied and pasted but cannot tell to which part is copied and which part is pasted. One oI the box parts in the signal is copied and then pasted in the other box. Also the two oval shape boxes tell some Ialse matching pairs which were also detected. Because total accuracy is not possible to achieve, we tried to detect the doctored part as much as possible and we succeed in almost all the tested audio clips. Now to deIine the accuracy oI the proposed method, Iirst, the perIormance and contribution oI the Ieatures that were used in the method will be explored We have used set oI seven Ieatures those are zero crossing rate(ZCR),Root mean square(RMS).Mel-Irequency cepstral coeIIicients(MFCC),Spectral Flatness Measure(SFM),Spectral Crest Factor(SCF),Power Spectral Density(PSD) coeIIicients` mean and Power Spectral Density coeIIicients` Standard Deviation. The requirement oI so many Ieatures used in the matching process is essential. Let L is the length oI the signal in which r is region which is copied and pasted in the signal, The resulting tampered signal length will be (Lr) and the region (L-r) is unmanipulated. The genuine region oI the signal retrieved by our method is called Ialse matching pair. The reason behind to use 7 Ieatures can be understand by analyzing the percentage oI Ialse matching pair retrieved 192 Elsevier Publications, 2013 Abhishek Verma, Vivek Kumar Singh and R.C. Tripathi RMS value and ZCR value only 70 RMS value, ZCR value,MFCC with threshold 45 RMS value, ZCR value,MFCC value with appropriate threshold value,PSD coeIIicients mean and slandered deviation value 30 SFM value and SCF value with appropriate thrsholding 20 All Ieatures used 8-10 5. Conclusion In this paper, the topic oI copy-move Iorgery is discussed in digital speech audio signals Ior the purpose oI authenticity by doing matching process. In the time-domain oI the signal, windowing over the signal is done and treating each window itselI as a signal. By using most suitable audio Ieatures, the method shows the region oI suspected Iorged part oI the signal with those matching pairs oI the window, which is returned by shiIt vector calculation. Those Ieatures which are known Ior their robustness like spectral Ilatness and spectral centroids were chosen. These Ieatures give us the chance to Iind the matching pairs within some threshold limit constraint. Except the time taken by the experiment, it deIinitely decides about the authenticity oI the signal. We are currently in the process oI developing new methods oI detecting copy-move Iorgery except this method, by using the beneIit oI the audio Ieatures. Acknowledgement: We are thankIul to IIIT Allahabad Ior providing us necessary resources in Speech, image and Language processing Lab. We are also thankIul to TCS RSP Ior their support Ior the work. References |1| Shrishail Math and R.C.Tripathi. Digital Forgeries: Problems and Challenges International Journal oI Computer Applications (0975 8887), Vol 5 No.12, August 2010. |2| Hanna Greige and Walid Karam.. Audio-Visual Biometrics and Forgery Advanced Biometric Technologies. Dr. Girija Chetty (Ed.), ISBN: 978-953-307-487-0, 2011. |3| V.K.Singh and ,R. C Tripathi. Fast and EIIicient Region Duplication Detection in Digital Images using Sub-Blocking Method International Journal of Advanced Science and Technology, Vol. 35, pp. 93-102, 2011. |4| Robert C. Maher. Audio Forensic Examination |Authenticity, enhancement, and Interpretation| IEEE signal processing magazine |84|, March 2009. |5| ShinIeng D. Lin and Tszan Wu. An Integrated Technique Ior Splicing and Copy-move Forgery Image Detection 4th International Congress on Image and Signal Processing, 2011. |6| Dalibor Mitrovic, Matthias Zeppelzauer and Christian Breiteneder. Features Ior Content-Based Audio Retrieval Advances in Computers, Vol.78, pp.71-150, 2010. |7| Jiirgen Herre, Eric Allamanche and Oliver Hellmuth. Robust Matching OI Audio Signals Using Spectral Flatness Features IEEE Workshop on Applications of Signal Processing to Audio and Acoustics 2001, 21-24 October 2001. |8| Rui Yang, Zhenhua Qu and Jiwu Huang Detecting Digital Audio Forgeries by Checking Frame OIIsets ACM 978-1-60558- 058,MM&Sec08,September 2223, 2008, |9| RaIal Korycki. Methods oI Time-Frequency Analysis in Authentication oI Digital Audio Recordings International Journal of Electronics and telecommunications, Vol. 56, September 2010, |10| Sevinc Bayram, Husrev Taha Sencar and Nasir Memon An EIIicient and Robust Method Ior Detecting Copy-Move Forgery IEEE, ICASSP 2009. |11| Bruce E.Koenig and Douglas S. Lacey. Forensic Authentication oI Digital Audio Recordings J. Audio Eng. Soc.,Vol. 57, No. 9, September 2009. |12| Swati Gupta. Current Developments and Future Trends in Audio Authentication IEEE Computer Society, 2012. 193 Elsevier Publications, 2013 Table 2. Comparison oI results on the basis oI diIIerent Ieatures Features used Percentage of false matching pair A assive pproach to etect opy-ove orgery in igital peech udio ignal |13| Christian Kraetzer, Andrea Oermann, Jana Dittmann and Andreas Lang. Digital Audio Forensics: A First Practical Evaluation on Microphone and Environment ClassiIication ACM 978-1-59593-857-2.September 2007. |14| J. S. GaroIolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S.Pallett, and N. L. Dahlgren. DARPA TIMIT acoustic-phonetic continuous speech corpus CD-ROM, 1993. National Institute of Standards andTechnology, NISTIR 4930. |15| ABC News. Did someone mess with mel gibson`s audio recordings? http://abcnews.go.com/Entertainment/ mel-gibsons-rants- messed/story? id11169736, July, 2010 194 Elsevier Publications, 2013 Index