Ryan Ye

How can we apply pre-trained transformers for generation of MIDI music files given input by a user?

Started Mar. 16, 2023

Abstract or project description

Music generation is an intriguing machine learning task due to the necessity of both short and long-range structure. With the increasing relevance of pre-trained language-based transformer models such as ChatGPT, we are interested in applying a similar transformer architecture to music generation. Transformers are unique in their ability to represent long-range structure, a clear advantage over recurrent neural networks. In this work, we will fine-tune pre-trained transformers for MIDI music generation. We will first obtain a database of MIDI files, using existing MIDI databases such as Lakh Pianoroll Dataset and the MAESTRO Dataset. Then, we will fine-tune a large transformer model for next-note prediction to generate new music given a user input of a short composition. Overall, in this project we leverage the power of transformers for generation of MIDI music given music input from a user.