Residual-enhanced transformer for affective multi-part music generation

Music generation demands modelling intricate multi-dimensional sequences while preserving structural coherence and emotional expressiveness. To address transformer's limitations in detail retention, multi-track efficiency, and affective integration, we propose residual enhanced affective music transformer (REAM) with three key innovations: 1) residual dense blocks establishing inter-layer skip connections to enhance feature reuse and maintain fine-grained musical textures; 2) emotion-aware rotary positional encoding that dynamically modulates note relationships based on target sentiment vectors; 3) lightweight residual modules enabling efficient parallel generation of multi-track compositions. Through systematic ablation studies and perceptual evaluations, REAM demonstrates superior performance in both objective reconstruction metrics and subjective musicality assessments. This framework bridges symbolic precision with affective depth, enabling computationally efficient generation of structurally coherent, emotionally controllable multi-instrument music compositions.