Analysis and automatic detection of double compressed audio signals

In thesis study, double compressed audio detection problem will be investigated. As a result of widespread usage of audio recording as evidence in courts, this addresses the problem of authenticity of a audio content. In order to tamper an audio file, attacker first needs to decompress it to waveform and then manipulate the waveform which is followed by recompressing the manipulated audio waveform into the original format. For this reason, double compressed audio detection is an important audio forensic problem. Previous studies addressing the double compressed audio detection problem with prior knowledge about encoding and decoding processes of codec. In this study, it is aimed at detecting double compressed audio with extracting general feature when there is no prior knowledge about codec. There are a few studies including deep learning approaches on double compressed audio detection problem. The proposed system is used for both feature extraction and classifier with deep learning methods. The system will be developed to be succesful against noise, channel and recording environment variations. The system will be invegstigated using different deep learning models and datasets detailed. The performance of the system will be developed on cross-database evaluation. One of the dataset including audio with compression history will be investigated on double compressed audio detection problem. In the first part of thesis, single and double compressed audio recordings will be generated with narrow band AMR codec. Audio spectrograms will be used as input of deep learning model. The system will be improved with different deep learning models and different loss function to try on better performance. There is no previous study addressing the double compressed audio detection problem including wide band AMR codec. The system will be developed double compressed audio compressed by wide band AMR codec.