Comparing LSTM and GANs in automatic music composition

In this work, composing classical music using Long Short-Term Memory (LSTM) models and Generative Adversarial Networks (GAN) was explored. The proposed thesis compares both state-of-the-art LSTM and GAN networks in classical music composition as there is a lack of research in comparing these deep learning models in music composition.

The dataset used in this work contains 227 Music Instrument Digital Interface (MIDI) songs consisting of 19 world-renowned classical composers including J.S Bach, L. Beethoven and W. Mozart. Furthermore, evaluation remains a research problem in automated music composition.

This study proposes an objective and subjective evaluation to compare both models. Evaluation metrics from computational music theory and a user study were conducted in which the study concluded that the LSTM network produced better results than GAN network in classical music composition.

Recieved the 2nd highest grade in the BSc program at 8.5/10