Towards the generation of melodic structure
Journal:
Proceedings of the Fourth International Workshop on Musical Metacreation
Description:
This research explores the generation of melodies through the creation of a probabilistic analytical model of melodies. Using a Natural Language Processing technique utilized for the automatic reduction of melodies: the Probabilistic Context-Free Grammar (PCFG), it is possible to reduce new incipit melodies, or to generate and embellish melodies. Automatic melodic reduction has been previously explored by means of a probabilistic grammar (Gilbert and Conklin 2007) (Abdallah and Gold 2014). However, these methods used unsupervised learning to estimate the probabilities for the grammar rules rather than a corpus-based evaluation. A treebank of analyses using the Generative Theory of Tonal Music (GTTM) exists (Hamanaka, Hirata, and Tojo 2007), which contains 300 Western tonal melodies and their corresponding melodic reductions in tree format. In this work, a new representation of the GTTM grammar is created using a higher-level representation based on intervals. Then, supervised learning is used to train a PCFG on the treebanks with different versions of the new data representation. The resulting model is evaluated on its ability to create accurate reduction trees, and its ability to generate melodies is subjectively explored. With this approach, each generated melody will not only be a sequence of pitches, but also a hierarchical structure containing the melodic reduction of the generated melody. Multiple data representations are tested, and example output reductions—and embellishments—are shown.