Part-invariant Model for Music Generation and Harmonization
Yujia Yan, Ethan Lustig, Joseph VanderStel, Zhiyao Duan · Zenodo (CERN European Organization for Nuclear Research) · 2018
Automatic music generation has been gaining more attention in recent years. Existing approaches, however, are mostly ad hoc to specific rhythmic structures or instrumentation layouts, and lack music-theoretic rigor in their evaluations. In this paper, we present a neural language (music) model that tries to model symbolic multi-part music. Our model is part-invariant, i.e., it can process/generate any part (voice) of a music score consisting of an arbitrary number of parts, using a single trained model. For better incorporating structural information of pitch spaces, we use a structured embedding matrix to encode multiple aspects of a pitch into a vector representation. The generation is performed by Gibbs Sampling. Meanwhile, our model directly generates note spellings to make outputs human-readable. We performed objective (grading) and subjective (listening) evaluations by recruiting music theorists to compare the outputs of our algorithm with those of music students on the task of bassline harmonization (a traditional pedagogical task). Our experiment shows that errors of our algorithm and students are differently distributed, and the range of ratings for generated pieces overlaps with students' to varying extents for our three provided basslines. This experiment suggests some future research directions.