Music generation using deep learning

Abstract

This research studies the generation and the results of a Text-to-Midi Transformer model, focusing on converting textual descriptions to functional and melodious sounds in MIDI format. A novel dataset on combinational music is used to enable the model to learn richer representations of music. This study uses an advanced music information framework to label music files, to improve accuracy on music captioning. An advanced text LLM is employed to process music captions, to improve understanding of their nuanced meanings. A user study is used to evaluate the hypothesis that a Text-to-Midi deep learning model can be designed to generate short, emotion-driven thematic music from textual prompts. The results verify that the model can generate music that is melodious, and able to capture the nuances of the textual prompt. A model that can produce melodious sounds better than previous implementations of other students while being able to capture emotional context from textual prompts is established.