A Preliminary Study on Models for AI-based Arrangement Generation

Abstract

This paper systematically reviews the current research status and developmental trajectory of artificial intelligence in the field of music arrangement. Artificial intelligence simulates the creative process through computational models and has undergone a transition from symbol-based rule-driven algorithms to data-driven deep learning paradigms. By learning statistical regularities from massive music datasets, it enables the generation of melody, harmony, and rhythm. However, the structural complexity and artistic logic of polyphonic music remain the core challenges for AI-based arrangement. Current mainstream approaches include autoregressive neural generative models, variational autoencoders (VAE), and generative adversarial networks (GAN). This paper focuses on the analysis of the GAN-based MuseGAN model and its evolutionary framework, the Leadsheet Generation model. By introducing functional representations (melody and harmony) as structured prior conditions, Leadsheet Generation constructs a conditional generation mechanism, significantly improving the quality of music in terms of tonal consistency, textural repetition patterns, and auditory logic. This provides a new path for human–computer interactive music creation that balances controllability and artistic expressiveness.