Toward Guided Musical Form: Structuring Generative Music with LLMs

Long-form music generation remains a challenging task for generative models, particularly in capturing coherent musical structure over extended durations. This article introduces a hierarchical approach that combines a large language model (LLM) with a transformer-based audio generator. The LLM designs the overall musical form and generates natural language prompts for each section, whereas the audio model synthesizes the corresponding music. To evaluate structural coherence, the current study proposes a new quantitative metric, in which the structural patterns of generated music are compared with those of human-composed pieces. The method is evaluated using both objective metrics and subjective listening tests. On analyzing the results, it appears that music with a more coherent and varied structure across temporal scales is produced. Additionally, a metaoptimization method is introduced, by which the prompt design process for the LLM is automated and generation quality is improved through iterative refinement.