Evaluation of state of the art for genre classification in large datasets

TitleEvaluation of state of the art for genre classification in large datasets
Publication TypeMaster Thesis
Year of Publication2018
AuthorsBajpai, V.
AbstractThe goal of this thesis is to evaluate state of the art methods for genre classification on some popular genre datasets and provide an alternate dataset for the music community to use for the task of genre classification. Genre classification has been one of the main classification tasks in the MIR community due to its direct use in auto tagging of songs by the research community and the music industry alike. Companies like Spotify, YouTube, SoundCloud find it essential to tag songs based on genres since it is an important way to sort songs and gauge the listening styles of the users. It is, therefore, essential to study the features and models which might help in better classifying the songs based on various genres. However, there aren’t many quality datasets which have songs which are publicly available, balanced in terms of the number of genre classes present in them, and clean audio with longer durations. This thesis is an attempt to create models with better feature selection, and creating a dataset leveraging the publicly available Jamendo audio set for the task of genre classification. The presented models and dataset creation methods are provided with the aim to create better generalizable audio dataset.
KeywordsGenre classification, Jamendo