Melody Infilling with User-Provided Structural Context

Considering composers usually write musical pieces in In recent years, machine learning techniques have been a hierarchical manner [25], we note that prompt-based conditioning widely applied to symbolic music generation. A large approaches have a strong limitation: they generate number of models attain sequential generation by accounting results with only consideration of local smoothness for only the past context, i.e., the generated music depends among the past context, future context, and result, without on only the preceding musical content [1-14]. While taking care of the overall musical form or structure of the sequential generation can find useful use cases, it does not music. For instance, a composer may like to write a song align with typical human compositional practices which in a musical form of ABA'B'. If we consider the concatenation can be non-sequential in nature. Musicians often write motifs of the segments corresponding to A and B (i.e., AB) or small pieces to get inspiration first, before working as the past context, and the segment corresponding to B' on the middle parts to connect them.