Following an existing score should be a lot easier than de novo polyphonic transcription. For example, you could almost just track the beat while ignoring all the pitches.
People will want to play the first few measures over and over, try something again after messing it up, stop and take it from the top and so on. If you turn the page too soon or scroll the music the user is playing off the screen, the customer reviews are going to look like Armageddon has come.
The only way I see that works understandably and well enough is to recognize when the last measure of the last staff is completed. Even that may be incompatible with some music, but it should work well enough in general.