Pop Music Midi Generation using Long Short-Term Memory based Recurrent Neural Network

Abstract

As music composition and technology grow, the pursuit for music creation with machine learning as a trans-human collaborator is a relatively new venture in the world of artificial intelligence. Various RNN models for music genres such as classical, folk, and jazz has been made but none specifically for pop music. The study aims to fill in this gap. A pop music dataset was created from Billboards' Hot 100 Year End Charts of 2021. The dataset was used to train LSTM-based RNN models which were then evaluated through loss. This revealed that the model with the lowest loss came from the ADAM optimizer making it the best choice with 200 epochs. The results of the chosen model were then evaluated through its perplexity and a listening test. With a low perplexity score, the model was deemed confident in creating novel midi samples. The listening test then tested the created midi samples, resulting in a good rating.