SYSTEM OF AUTOMATIC SEGMENTATION OF PAUSES IN PHONOGRAMS ON THE BASIS OF NEURON NETWORKS OF THE DEEP LEARNING

Abstract

The use of neuron networks of the deep learning for the construction of tool for realization of examinations of materials and apparatus of the digital audio recording allows to solve the «frigging» problem of such examination — problem of exposure of tracks of editing in digital phonograms. These networks provide high probability of exposure of such tracks in the pauses of speech information writtenin on a phonogram. Before man-hunting of tracks of editing in the investigated phonogram it is necessary to distinguish pauses (to perform its segmentation), and tool built on the basis of neuron networks of the deep learning, requires its work to be done in automatic mode. The basic requirement of automatic segmentation is high efficiency of selection of pauses in the conditions of permanent change of level of noises in phonograms. It is determined by probability of errors of І and ІІ kinds. It is offered on the basis of neuron networks of the deep learning to create CAS of segmentation of phonograms, possessing high efficiency of selection of pauses in speech information. Thus the system must be independent of level of noises in every concrete pause, and also language, context and announcer, whose speech is fixed in a phonogram. It is suggested to examine pauses as one of the types of voice information, which characteristics differ from characteristics of speech information fixed in a phonogram. For educating of such network it was required to create the primary base of these sounds and pauses. On its basis three arrays of the data, intended for learning, testing and determination of the crooked errors of І and ІІ kinds, are created. After learning and testing the system passed verification on the real phonograms. As a result taking into account some features of speech on the neuron networks of deep learning there has been built the system providing effective segmentation of pauses in phonograms in the automatics mode. The obtained results suit examination that is conformed by given curves over of errors of І and ІІ kinds.