Composer as Constrainer in Model-Led Piano Sheet Music Generation

We describe a generative music system designed to produce expressive MIDI music transcribed into scores for the piano. The approach entails the constrained application of a transformer neural model which generates expressive MIDI music. The constraints enable an interpretation of the transformer’s output tokens so that they can be captured as notes on sheet music. Additional functionality affords more control over the form, difficulty and playability of the music, enabling composers to direct the music generation by specifying a series of constraints. Producing the best results requires a model-led approach whereby the notes the transformer produces solves a constraint satisfaction problem (CSP), but the nature of the solution also dictates which CSP to solve next. We assess progress via feedback from two pianists about nine generated pieces, in terms of expressivity, quality, likeability and playability of the sheet music, and show that the pieces are in general well-received.