Analysis and Comparison of Algorithmic Composition Using Transformer-Based Models

Shaozhi Pi

2024

Abstract

As a matter of fact, transformers have revolutionized generative music by overcoming the limitations of earlier models (e.g., RNNs) in recent years, which struggled with long-term dependencies. With this in mind, this paper explores and compares four transformer-based models, i.e., Transformer-VAE, Multitrack Music Transformer, MuseGAN as well as Pop Music Transformer. To be specific, the Transformer-VAE offers hierarchical control for generating coherent long-term compositions. In addition, the Multitrack Music Transformer excels in real-time multitrack music generation with efficient memory use. At the same time, MuseGAN supports human-AI collaboration by generating multitrack music based on user input, while Pop Music Transformer focuses on rhythmic and harmonic structures, making it ideal for pop genres. According to the analysis, despite their strengths, these models face computational complexity, limited genre adaptability, and synchronization issues. Prospective advancements, including reinforcement learning and multimodal integration, are expected to enhance creative flexibility and emotional expressiveness in AI-generated music.

Paper Citation

in Harvard Style

Pi S. (2024). Analysis and Comparison of Algorithmic Composition Using Transformer-Based Models. In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-754-2, SciTePress, pages 184-191. DOI: 10.5220/0013512300004619

in Bibtex Style

@conference{daml24,
author={Shaozhi Pi},
title={Analysis and Comparison of Algorithmic Composition Using Transformer-Based Models},
booktitle={Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2024},
pages={184-191},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013512300004619},
isbn={978-989-758-754-2},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Analysis and Machine Learning - Volume 1: DAML
TI - Analysis and Comparison of Algorithmic Composition Using Transformer-Based Models
SN - 978-989-758-754-2
AU - Pi S.
PY - 2024
SP - 184
EP - 191
DO - 10.5220/0013512300004619
PB - SciTePress