Projects in expressive music generation, musical representations, and computational creativity.
A unified representation space for piano performance generation and assessment.
A collaborative framework that learns representations of piano performances across different levels of expertise and recording contexts. It connects score-conditioned generation with score–performance representation learning. An ISMIR 2026 main conference paper, to appear, and a Best Paper Award nominee.
Pipa-centric data and heterogeneous graphs for Nanyin ensemble generation.
A collaborative project modeling heterophonic interaction in Nanyin ensembles through a heterogeneous graph neural network. I contributed to dataset construction and model training. A CSMT 2025 conference paper bringing oral-tradition knowledge into symbolic music generation.
Adding expressive ornaments to MIDI with music-aware tokenization and a dedicated Transformer.
A Transformer-based system for expressive MIDI ornament generation using PerTok tokenization and syntax-constrained sampling. An ISMIR 2025 Late-Breaking/Demo contribution on representing expressive details in symbolic music.
An expressive piano rendering model selected as a RenCon 2025 finalist.
An expressive piano rendering model developed for the RenCon 2025 contest at ISMIR. Selected as a finalist, its renderings were played on a Yamaha Disklavier during the blind evaluation of expressive quality.
An interactive performance system using timbre analysis and granular synthesis.
A visual and sonic exploration of timbre spaces using spiral trajectories, real-time MFCC analysis, k-nearest-neighbor interpolation, and granular synthesis. The interface displays algorithmic states and sonic transformations as part of the performance.
A preliminary study of dynamic audio chunking for irregular music generation.
A preliminary study proposing dynamic audio chunking and variable-length token representations for non-isochronous music generation. A contribution to the LLM4Music workshop at ISMIR 2025.
Examining interpolation failures in MusicVAE’s latent space.
A study of perceptual discontinuities in MusicVAE interpolation through failure-case analysis and latent-space visualization. Published in the proceedings of the third Workshop on Human-Centric Music Information Research at ISMIR 2025.
Chordless, narrative-driven heterophonic multi-track MIDI generation.
A project exploring the relationship between narrative and the heterophonic texture of Nanyin through multi-track MIDI generation. Presented as a poster at the New in ML workshop at ICML 2025.