Latent Phrase-Aware Generative Modeling forExpressive Symbolic Audio Synthesis

This paper presents a generative AI frameworkfor producing structured symbolic sequences with fine-grainedexpressive control. The approach introduces a compact tokenrepresentation combined with phrase-aware latent alignment tosupport coherent generation across variable-length segments. Byintegrating sequence-level regularization directly into attention,the model balances structural consistency and diversity withoutrelying on explicit post-processing constraints. Empirical analysisshows that the method maintains stable distributional behavioracross expressive dimensions, highlighting its suitability forcontrollable symbolic generation tasks.