Application of recent methods of audio signals' feature extraction to classification of studio recorded and live performed tracks.

The aim of this thesis is a comparative analysis of various approaches to the classification of music recordings in studio and concert recording forms. The approaches are divided into two groups: the so-called "classical" methods (implemented as baseline methods) and deep learning methods ("modern" methods). In this paper, the emphasis is placed on deep learning methods. Various architectures of deep learning models have been implemented and investigated. The architecture of such a deep model consists of a frontend and backend-parts. The effect of using different front-end and backend parts on the result according to selected quality metrics was first investigated. The influence of the choice of audio recording segment on the final result was also investigated. It was shown that the most relevant segment is the initial segment. Last but not least, it was shown that the use of augmentations is preferential for this task in the general case.