Visual-based Musical Data Representation for Composer Classification

Abstract

Automated classification for musical genres and composers is an artificial intelligence research challenge insofar as music lacks a rigidly defined structure and may result in varied interpretations by individuals. This research collected acoustic features from a sizable musical database to create an image dataset for formulating a classification model. Each image was constructed by combining pitch, temporal index length, and additional incorporated features of velocity, onset, duration, and a combination of the three. Incorporated features underwent Sigmoid scaling, creating a novel visual-based music representation. A deep learning framework, fast.ai, was used as the primary classification instrument for generated images. The results were that using velocity solely as an incorporated feature provides optimal performance, with an F1-score of 0.85 using the ResN$e$t34 model. These findings offer preliminary insight into composer classification for heightening understanding of music composer signature characterizations.