Method of identification of attributes of phonogram digital editing using neural networks of deep learning

In the recent decade, the models on the neural networks of deep learning are effectively used for solving many actual tasks requiring large data arrays processing. The important task of expertise of materials and apparatus of the digital audio recording belongs to them, i.e., automatic identification of tracks of digital processing (tracks of digital editing) of phonograms. Editing of phonograms is produced in pauses of the speech information. Therefore, the search for traces of digital processing is reduced to finding attributes of such editing in pauses fixed on phonograms, and it is necessary to attribute this process to the task of binary classification. The complexity of the construction of such system consists in, first, attributes of such editing are extremely small and, second, their selection from the signals of pauses by the known classical processing methods is very problematic. The basic requirement to the expert tool is the ability to provide a selection and obvious demonstration of signs of editing. Thus, an expert must be convinced in the reliability of the results of expertise. Therefore, the generally accepted impossibility of establishing a connection between both signals on an input with the results obtained on the output of the applied model and occurring processes is the major factor of influence on further development of the systems of court expertise on neural networks. Authors suppose that for some tasks of binary classification, in particular, tasks of identification of digital editing of phonograms, such possibility exists. The objective of the research is a method of identification of features of digital editing of phonograms holding the expertise requirements, based on the use of a neural network of deep learning. The method of identification of traces of editing is proposed and considered for pauses between signals of the speech information with the use of a neural network of deep learning. It is suggested to identify pauses of speech with traces of editing by the binary classification in a network. Supplementary processing of the modeling results provides their graphical interpretation, that enables the selection of fragments of pauses with a high degree of probability of correct classification in the separate array. This provides the possibility of the creation of CAS of identification of traces of editing in digital phonograms. © 2020 by Begell House Inc.