Making music directly as sound is difficult: a system must get tiny details right while keeping rhythm and longer musical structure coherent.
The work splits the task into parts: one model turns recordings into note instructions, another creates note sequences, and a third turns those notes into sound. The note format, MIDI (Musical Instrument Digital Interface), makes the music easier to inspect and edit.
The released collection contains 172 hours of piano recordings, with note labels aligned to the audio within about 3 milliseconds. Together, the data and models support music generation with structure that people can understand and manipulate.