STARR: Similarity-based Transfers for Automatic Rock Rhythms

During the composition of a piece of music, composers often focus on the crafting of melody and harmonic lines rather than a percussive track that may accompany them. For amateur composers, creating a drum track by hand can be both tedious and frustrating, due to their lack of skill and experience with the task. In recent years, computer software tools have been created to help automate the creation of drum tracks, but most often they operate by simply allowing the composer to choose one of a set of drum patterns, and continually repeating that pattern for the duration of the song. As one might expect, this technique often results in aesthetically displeasing drum tracks, as their lack of variation quickly leads to listener boredom. The primary solution to this problem used by human drummers is the insertion of drum fills percussive elements that differ from the current drum pattern, intended to add energy and excitement to the performance of a musical piece by promoting listener interest. Unfortunately, the task of choosing both when to play a drum fill and what fill to play is non-trivial; the talents of human drummers are often rated by their ability to make these choices well. In this paper, I propose a novel solution to the problem of automatically choosing where in a melody-accompanying drum track to place drum fills; this problem is henceforth referred to as the Drum Fill Placement problem. Using a database of human-crafted melodies and accompanying drum tracks, a machine learning algorithm is trained to identify and exploit a relationship between melody rhythms and the occurrence of drum fills. Given a set of test melodies, the resulting drum fill placements are evaluated by comparison to hidden, human-crafted drum tracks designed to accompany each melody. The remainder of this section introduces a set of terminology needed for discussing the Drum Fill Placement problem, gives a brief background concerning the choice of music representation used, and formally states the goals of this work. Following this, the remainder of the paper is organized as follows: first, some specific refinements to the Drum Fill Placement problem as it is treated in this work are defined; second, a set of related work is discussed, with a focus on how it might be useful in solving the Drum Fill Placement problem; third, a novel algorithm designed to solve the Drum Fill Placement problem is introduced, titled STARR (Similarity-based Transfers for Automatic Rock Rhythms); fourth, a set of experiments designed to achieve the goals of this project are described, and results are given; fifth, the experimental results are discussed; finally, future work is suggested, and conclusions are drawn.