TASTE: Taxonomy-Aware Symbolic Piano Generation from Textual Description

Text-conditioned symbolic music generation lets users steer editable music with language, yet two questions remain underexplored: what a conditioning description should contain, and how data for such descriptions should be collected and used for training. We address both questions with TASTE, a taxonomy-aware framework for symbolic music generation with main focus on piano. First, we introduce a taxonomy that separates finite-vocabulary tags (key, tempo, emotion, genre, accompaniment, and performance style), a section plan for temporal form, and a free-form imagery description of the scene or atmosphere a piece evokes. Each part of a description can then be annotated, learned, and evaluated on its own terms. Second, we build a multi-tier annotation pipeline that draws each condition from suitable evidence and propagates annotations from 10K directly paired and 51K curated public performances to 371K weakly labeled MIDI files. Finally, we design a staged training recipe that moves through these tiers from broad weak supervision to direct pairs, and instantiate it in both an autoregressive decoder and a latent diffusion transformer. Both objective and subjective evaluations demonstrate substantial improvement on condition alignment at almost all fields over caption-conditioned baselines while maintaining competitive music quality. We will release the annotated datasets.