Structured Music Transformer: Structured conditional music generation based on stylistic clustering using Transformer

Recent advancements in music generation research have significantly progressed the field. However, a prevalent issue among current models is their tendency to overlook music’s intrinsic structure, leading to compositions that lack coherence and thematic consistency. This paper proposes a novel approach to structured music generation that addresses these challenges by considering music structure as styles of different subsegments. The approach enhances the deep structural aspects of generative music by deriving embedding vectors from pre-trained contrastive models and clustering them into discrete structured tokens for generative conditions. Furthermore, this study introduces a self-similarity-based evaluation method to assess the structural features of the generated music. Experimental results indicate that the Structured Music Transformer (SMT) significantly outperforms contemporary models in producing compositions that are stylistically coherent and structurally robust, as demonstrated across two diverse datasets.