Source Recording Device Identification In The Presence Of Additive Noise

Digital audio,as a common digital media carrier in everyday life,is often provided to the court as evidence in court cases.Therefore,verifying the source,authenticity,and integrity of digital audio is crucial.The authenticity of the source recording device for audio data is one of the hot issues of digital audio forensics technology and has achieved great results.However,the subjects studied are almost speeches recorded in a quiet environment,and rarely consider the situation in a noisy environment.It is more practical to identify the source recording device identification in the noisy environment and cell-phones have become the most popular recording device.Therefore,this paper studies the source cell-phone identification method under additive noise.The research work has been carried out mainly from the following three aspects:Firstly,for the construction of the database,clean and noisy speech databases containing 24 models cell-phones of 7 brands were recorded,namely CKC-SD? TIMIT-RD clean speech databases,CKC-SD?TIMIT-RD noisy speech databases.Based on the constructed speech databases,the noise robustness of two existing source recognition algorithms with excellent performance is detected.The experimental results show that although the current recognition algorithm has good recognition performance for clean speeches,the noise robustness is poor.The source cell-phone identification algorithms in a quiet environment cannot be generalized to noise conditions.Secondly,we specifically analyze the difference in the device of different brands of cell-phones and different models of cell phones from the same brand embedded in the speech by studying the spectrograms of the speeches.The effects of environmental noise on device differences and the effects of denoising methods on device differences have also been studied.The above research results lay the foundation for the source cell-phone identification algorithm in the noise environment.Thirdly,according to the difference of different devices in the frequency domain,it is found that the key to the source cell-phone identification lies in the differentiation of different models of cell-phone of the same brand.Therefore,this paper proposes a source cell-phone identification algorithm based on the constant Q transform(CQT)and multi-scene training.The experimental results show that the algorithm maintains the distinguishing performance of different brands of cell-phones,and also improves the recognition rate of cell-phones within the same brand.The recognition accuracy of clean speeches is 97.08.%?99.29%,respectively on the CKC-SD and TIMIT-RD databases.For noisy speeches,the average recognition rate exceeds 90%,which greatly improves the recognition ability of the algorithm for noisy speeches.Finally,by analyzing the difference information of different device difference in different frequency domains,a source cell-phone recognition algorithm based on fusion feature and convolutional neural network(CNN)is proposed.The experimental results show that the recognition rate of the algorithm,which uses different time-frequency transform methods to extract the fusion features of spectral distribution feature of short-time Fourier transform(STFTSDF),Mel Frequency Cepstral Coefficient(MFCC)and spectral distribution feature of constant Q transform(CQTSDF)as the device distinguishing feature,is 94.06% and 92.44% on the CKC-SD and TIMIT-RD noisy databases,respectively.In addition,through the differences in device characteristics and a large number of experiments,the input selection of CNN and the principles of CNN framework design for source cell-phone forensics are explored.