Decomposer Recovers Music Programs from Symbolic MIDI Data

Musical performance involves executing high-level instructions, but extracting these instructions from a performance is a complex inverse problem. This research introduces Decomposer, a post-training framework designed for symbolic music decompilation. Its goal is to recover editable music programs from symbolic music, specifically converting MIDI input into Strudel, a music programming language, such that the generated program can reconstruct the original MIDI. The framework tackles two main hurdles: the scarcity of paired MIDI-code data for Strudel and the risk of generating unreadable, note-by-note transliterations if only MIDI reconstruction is optimized. Decomposer addresses these by first creating "Strudel-Synth," a synthetic corpus of paired Strudel programs and MIDI, used for supervised fine-tuning. Subsequently, it refines the model using reinforcement learning on unpaired MIDI, optimizing for both faithful MIDI reconstruction and the readability of the generated code. Evaluations show Decomposer significantly outperforms closed-source LLMs in MIDI reconstruction while producing more readable and diverse code than heuristic converters.