Controllable music generation for fluid curriculum

dexmo There is a growing trend to teach playing an instrument such as
a piano using an automated learning system. The main challenge of such approaches is to find a suitable practice schedule for the learner. In an optimal scenario a system should be capable of using a meticulous performance tracking in order to select practice strategies and pieces that optimally suits to the level and motivation of the learner. Our system uses a multimodal parameter space characterizing music pieces [1] by their attributes. To this end, a Gaussian Process model is responsible for picking a suitable practice parameterization. To complement this approach, a controllable AI music generator that we envision generates pieces based on this very precise parameteric task description, consisting of rhythm and pitch entropy, note ranges, number of hands, parallelization, bpm, etc. Such an approach enables us to use optimal pieces for learner practice, extending the set of existing pieces that come from a typical rigid curriculum structure.