FLAGNet : Feature Label based Automatic Generation Network for symbolic music

SeongHyeon Go

Keywords: cGAN, RNN, MIDI generation, music

Abstract: The technology for automatic music generation has been very actively studied in recent years. However, almost in these studies, handling domain knowledge of music was omitted or considered a difficult task. In particular, research that analyzes and utilizes the characteristics of each bar of music is very rare, even though it is essential in the human composition. We propose a model that generate music with musical characteristics of bars by conditional generative adversarial network, and analyze the good combination of the sequence of which characterized bars for symbolic-domain music generation by Recurrent Neural Network with Long short term memory layer. Also, by analyzing symbolic music data as image-like based on relational pitch approach, it increases the utilization of the data set with arbitrary chord scales and enables the use of generational results extensively. The resulting model FLAGNet generates music with the understanding of musical domain knowledge while handling inputs like minimum unit of note, length of music, chart scales, and chord condition.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: Creative and artistic music generator with understanding musical domain knowledge but a bit unstable.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=ITB4HTwae-