In the Euterpea for Haskell documentation it mentions that you're still using virtual instrument suites to sample the music, which brings you right back to size vs. quality. Also you would be working in Haskell which isn't exactly the most popular language, that will limit the number of resources available to you.

I mentioned Euterpea because it's a fairly straightforward example of the architecture I would use for a score manipulation system, but the same concepts can be (and have been) easily implemented in any language. For example Python has Mingus, Java has JFugue, and so on. Most of these systems generate or export MIDI because it's useful for a musician, but better integration with a synthesizer is certainly possible.