I'd have a many to many relationship between instrument/musician and track (an album consists of tracks, of course; didn't want to go into such detail because your "problem" seemed to be the boxed set - album relationship). Wouldn't make sense otherwise (database normalization) because it would be a hassle to even do simple queries like (pseudocode follows) SELECT track, artist.first_name, artist.last_name FROM [tables and join bla bla] WHERE instrument = 'kbd'. Is this a test?
Yes - I agree it would be a hassle - and not really produce much of an interesting result. That's why I went for putting all instruments in one field. When I get it finished, if ever, I'll be able to pick out all albums with a trombonist on them, or a named trombonist, and that'll do me.
Actually I meant the opposite. If I'd care to design a database for my music collection (I don't because I'm too lazy to enter data) I would not put all instruments in one field but have a record for each instrument a musician plays on a track. If significant performance backdraws aren't to be expected I go for higher granularity. One never knows what kind of queries Rooster Ties comes up with (like "on how many different instruments do you have artist x on your records where the instrument is not a piano?").