Music Genre Classification: A N-Gram Based Musicological Approach

Abstract

Digitalization of music has grown deep into people's daily life. Derived services of digital music, such as recommendation systems and similarity test, then become essential for online services and marketing essentials. As a building block of these systems, music genre classification is necessary to support all these services. Previously, researchers mostly focused on low-level features, few of them viewed this problem from a more interpretable way, i.e., a musicological approach. This creates the problem that intermediate stages of the classification process are hardly interpretable, not much of music professionals' domain knowledge was therefore useful in the process. This paper approaches genre classification in a musicological way. The proposed method takes into consideration the high-level features that have clear musical meanings, so that music professionals would find the classification results interpretable. To examine more musicological elements other than additional statistical information, we use a dataset of only symbolic piano works, including more than 200 records of classical, jazz, and ragtime music. Feature extraction and n-gram text classification algorithm are performed. The proposed method proves its concept with experimental results achieving the prediction accuracy averaged above 90%, and with a peak of 98%. We believe that this novel method opens a door to allow music professional to contribute their expert knowledge meaningfully in the music genre classification process, the proposed approach would contribute significantly for future music classification and recommendation systems.