Examining the synergistic integration of Artificial Intelligence and Music Information Retrieval for optimizing music generative model with similarity and esthetical perspective

Abstract

With the ongoing evolution of AI techniques, these technologies are actively pursued, resulting in significant achievements across various fields, including music generation. However, the generation process is still undergoing development and faces challenges, particularly in the realm of music. Since music is intangible and aesthetic, research in this area has taken diverse approaches. One significant challenge lies in acquiring data for the training process. Sound, being complex data, is currently captured through spectrum and MIDI data. While MIDI data remains more popular in music generation, limitations persist in obtaining data from musical instruments with and without MIDI systems. This research examines not only how the training data is obtained from the MIDI system but also the development of a generative system using an LSTM neural network and optimization based on similarity calculations via cosine similarity. Finally, this research also investigates how creativity can be achieved with the developed generative system using both objective and subjective evaluation methods.