Improvements in Texturizing Data for Automated Roman Numeral Analysis With Neural Networks
by Arisa Okamura
Abstract – Music performers, both professional and amateur, can benefit from Roman numeral analysis (RMA) of a musical composition, which provides the context and the significance behind each note. Recently, neural networks have been used to perform automated RMA. One such network, termed AugmentedNet, is trained with data augmentation by “texturizing” or modifying augmented input training data. This paper introduces two new texturizations, “arpeggios” and “auxiliary notes”, and shows that the combination of these texturizations statistically significantly improves performance upon AugmentedNet. Moreover, the model trained with these new texturizations produces analyses that demonstrate fundamentally deeper musical understanding.