Recognizing music characteristics using machine learning

Rad se bavi primjenom strojnog učenja na problem procjene emocionalnog doživljaja glazbe. Razmatrane su različite psihološke teorije emocija te psihološka perspektiva odnosa glazbe i emocija. U sklopu rada prikupljen je skup podataka koji glazbenim isječcima dodjeljuje ocjenu emocije. Na tom skupu podatka treniran je model dubokog učenja koji nije uspio postići željene rezultate, vjerojatno zbog male veličine skupa podataka. Razmotreni su različiti načini obrade digitalnih zvučnih signala i na što je potrebno obratiti pozornost prilikom pripreme zvučnih podatka za strojno učenje. Implementirana je aplikacija za korištenje modela koja reproducira zvučnu datoteku, vizualizirajući zvučni signal te pomoću modela procjenjuje prenesenu emociju reproducirane glazbe.The paper deals with the application of machine learning to the problem of assessing the emotional experience of music. Various psychological theories of emotion were examined and the psychological perspective of the relationship between music and emotions was given a closer look. As part of the work, a dataset was collected that assigns an emotion rating to music clips. A deep learning model was trained on that dataset and failed to achieve the desired results, probably due to the small size of the dataset. Different ways of processing digital sound signals are discussed and what needs to be paid attention to when preparing sound data for machine learning. An application has been implemented for using the model. It reproduces a sound file, visualizing the sound signal and uses the model to evaluate the conveyed emotion of the reproduced music