IDENTIFICATION OF RECORDINGS OF SOUND SIGNALS USING NEURAL NETWORKS

This paper considers the actual problem of calculating the percentage of similarity of recorded audio signals for the analysis and classification of audio data. The description of the algorithm for obtaining a special digital fingerprint of an audio recording (audio file) is given. The technique of forming a vector of features and a comparison procedure have been tested, which allows comparing the recordings of sound signals from similar sources and establishing their identity with an accuracy of 61% –65%, as well as correctly identifying obviously different sound signals. A comparative analysis of the calculated bit sequences obtained using the CREPE pitch tracker, which is a trained neural network, has been performed. In the case of audio signals, for the extraction of features from the recorded signals, it is taken into account that the signals reflect not only the features of the control object, but also distortions, random processes of various nature and other interfering factors.