Automatic Melodic Reduction Using a Supervised Probabilistic Context-Free Grammar

This research explores a Natural Language Processing technique utilized for the automatic reduction of melodies: the Probabilistic Context-Free Grammar (PCFG). Automatic melodic reduction was previously explored by means of a probabilistic grammar [11] [1]. However, each of these methods used unsupervised learning to estimate the probabilities for the grammar rules, and thus a corpusbased evaluation was not performed. A dataset of analyses using the Generative Theory of Tonal Music (GTTM) exists [13], which contains 300 Western tonal melodies and their corresponding melodic reductions in tree format. In this work, supervised learning is used to train a PCFG for the task of melodic reduction, using the tree analyses provided by the GTTM dataset. The resulting model is evaluated on its ability to create accurate reduction trees, based on a node-by-node comparison with ground-truth trees. Multiple data representations are explored, and example output reductions are shown. Motivations for performing melodic reduction include melodic identification and similarity, efficient storage of melodies, automatic composition, variation matching, and automatic harmonic analysis.