Song Content Labeling using an Incremental Learning Approach

Abstract

Multimedia Data Content Labeling has been a research area for a relatively long period. Researchers have spent a substantial amount of time focusing on classifying video sequences, music genres, and artists� classification. Artist tagging is prevalent in this field, but it is addressed chiefly to Western music. This thesis mainly focuses on the artist classification and tagging of the Indian songs, specifically in the Hindi language. Looking at the increasing amount of data being recorded by the Hindi song industry, old methods are not efficient enough for the artist tagging. Incremental learning has been used quite widely now in multimedia data content labeling. The proposed algorithm is developed for music vocal and multi-artist classifications using an incremental learning approach on the Hindi songs dataset. This incremental learning approach showed significantly good results. Cross-Language testing is also performed for Vocal/Music classification to check if the model generalizes over other language songs. Moreover, there are not enough datasets for performing such tasks for Hindi songs. Thus, a novel dataset is also introduced, containing window-based information labeled for each song for each artist. The dataset is primarily designed for tagging around twenty well-known Indian artists. This proposed method achieved remarkable accuracy of 83.6% and 55% for music-vocal and multi-artist classification on the test set, respectively.