Toward Interactive Music Generation: A Position Paper

Music generation using deep learning has received considerable attention in recent years. Researchers have developed various generative models capable of cloning musical conventions, comprehending the corpora, and generating new samples based on outcome. Although generated by these are persuasive, they often lack structure creativity. Moreover, a vanilla end-to-end approach, which deals with all levels music representation at once, does not offer human-level control interaction during process, leading to constrained results. Indeed, creation is recurrent process that follows some principles musician, where features reused or adapted. On other hand, piece adheres style, breaking down into precise concepts timbre performance composition coherency between aspects. Here, we study analyze current advances through different criteria. We discuss shortcomings limitations regarding interactivity adaptability. Finally, draw potential future research direction addressing multi-agent systems reinforcement algorithms alleviate limitations.