The Anticipatory Music Transformer (AMT) is a model that generates symbolic music (i.e. score generation such as MIDI) given user inputs, such as generating an accompaniment for a melody. Its name stems from the fact that during generation it anticipates user inputs in the near future, with the idea that better predictions are made with this additional context. While the original AMT (Thickstun et al.) is able to generate music, it only takes into account the duration and pitch of notes, but not the expressive aspects such as dynamics.
The aim of the project is to continue this development and evaluate its performance, with the goal that such a model can be deployed onto a Disklavier (an acoustic piano with recording and playback functionality) to enable live accompaniment generation.

As a Computer Science student who is currently learning about the quick developments in AI and a pianist myself, I am curious how AI can be used to collaborate with pianists, and wish to see the strengths and weaknesses of different models that accomplish this task. Unlike many other applications of AI, there is a large creative focus to music and seeing how the current models may be able to mimic this creativity interests me.