A chord-controlled transformer for controllable and coherent music generation

AI music generation faces significant challenges in musical coherence and user controllability, particularly in adhering to specific harmonic structures like chord progressions. While recent large-scale models have shown increasing promise, achieving long-term structural integrity remains a persistent challenge, particularly for models of a more moderate scale that aim for computational efficiency. Many existing end-to-end approaches still struggle to effectively integrate musical theory, often resulting in outputs with harmonic inconsistencies. To address this, we propose the Chord-Transformer, a novel architecture designed for chord-conditioned symbolic music generation. The framework features two main contributions: (1) a dynamic programming algorithm based on an energy function to extract the most salient chord progression from a given sequence, providing a robust high-level semantic guide; and (2) a parallel fusion architecture within the Transformer decoder that synergistically combines chord cross-attention with music self-attention. This mechanism, enhanced by a chord-aligned positional encoding, allows the model to maintain both local musical flow and global harmonic coherence. We conduct extensive experiments using the Pop909 and LMD datasets. Results from both objective metrics and a subjective user study show that the Chord-Transformer significantly outperforms existing mainstream models in controllability, structural rationality, and overall musicality.