Large Language Models for Symbolic Music Understanding

This thesis studies how Large Language Models can help computers “understand” symbolic music, such as MIDI, rather than only generating music. It first shows that when music is converted into text, the model can describe it but the descriptions are often vague and inconsistent. It then builds a scalable way to link symbolic music to meaningful labels and descriptions using online context. Finally, it trains a model that reads symbolic music directly and produces better captions and answers than a text-only approach. This work supports future tools for music analysis, search, education, and creative assistance.