So long as those probabilities are aware of context (style, form of this piece, location in the form, etc.). I've played around with Markov models a little, but I don't think they would quite cut it. This problem is probably slightly less difficult than generating prose from scratch, but it's a similar type of difficulty.

Getting solid curated input data would also be tricky, unless there's some big and well-annotated library of midi recordings sitting around somewhere.

For building a good app today, your best bet is almost certainly to use existing sheet music, pieces and etudes in the given style.