Pop Music Hook Generation Using Neural Networks

Abstract

While deep learning has recently gained popularity within a variety of disciplines in both industry and academia, applications of deep learning to music generation problems are comparatively less common. In this paper, I describe the process of designing from scratch a novel deep learning-based system for generating eight-bar-long pop hooks. I adopt an iterative approach where I design one model, analyze the merits and weaknesses of its output from both musical and technical perspectives, then use these to inform design decisions for subsequent models. My final model uses a recurrent neural network architecture and a generation strategy that feeds rhythmic "positional" encoding into the model along with standard event-based sequential data, which greatly improves its musical output. To verify that the model does not plagiarize, I rank the melodic similarity of the output to the rest of the dataset using a sequence similarity algorithm. Finally, I create an original dance-pop track, "Will You Learn", using a slightly modified version of an output example from my final model as the song's hook. The purpose of producing the song is to demonstrate the most likely use case for the model, which is to help kick-start musical inspiration for songwriters and/or producers. While my final model still has limitations, its musical quality is vastly improved from my first model and is at a level where an artist could potentially use it as a compositional tool.