Data interface implementation for music information retrieval

Abstract

Music elements are used as features for music information retrieval (MIR). However, while music elements are universal in music theory, there is no generic tool to extract them for modeling. Even worse, using music elements defined in theory directly as features require domain knowledge. This thesis presents a data interface that transforms music elements into basic music notation like pitch and duration, which lower the domain knowledge requirement. The interface converts between popular music file formats MIDI and XML and interchangeable format JSON, for a generic usage for music research. As an experiment, two datasets are created; one with 1613 classical songs in various epochs with 1.32 error rate and one with 128469 songs in various genres with 5.39% error rate.