Turkish Music Genre Classification using Audio and Lyrics Features

Music information retrieval (mir) has become a popular research area inrecent years. in this context, researchers have developed music information systems to findsolutions for such major problems as automatic playlist creation, hit song detection, andmusic genre or mood classification. meta-data information, lyrics, or melodic content ofmusic are used as feature resource in previous works. however, lyrics do not often used inmir systems and the number of works in this field is not enough especially for turkish. inthis paper, firstly, we have extended our previously created turkish mir (tmir) dataset,which comprises of turkish lyrics, by including the audio file of each song. secondly, wehave investigated the effect of using audio and textual features together or separately onautomatic music genre classification (mgc). we have extracted textual features fromlyrics using different feature extraction models such as word2vec and traditional bag ofwords. we have conducted our experiments on support vector machine (svm) algorithmand analysed the impact of feature selection and different feature groups on mgc. wehave considered lyrics based mgc as a text classification task and also investigated theeffect of term weighting method. experimental results show that textual features can alsobe effective as well as audio features for turkish mgc, especially when a supervised termweighting method is employed. we have achieved the highest success rate as 99,12% byusing both audio and textual features together.