Information Redundancy in Constructing Systems for Audio Signal Examination on Deep Learning Neural Networks

Preliminary signal processing methods used to create new tools to examine materials and digital sound recording means are described. It is shown that using information redundancy when creating a training base for deep learning neural networks used for such examination increases speaker identification efficiency based on voice characteristic parameters by about 15%. It is shown that the proposed processing methods enable speaker identification based on phonograms that are 1 second long.