I don't use Roman Rooms for 3x3x3 unless I'm doing multi BLD. And even then only when doing more than 2 cubes. And I've only tried that once - I tried 3 cubes a week ago and got the first 2 right and the third was off by a corner 5-cycle because I forgot to do it (as soon as I removed the blindfold I realized in horror that I had forgotten to do the CP, and I still had perfect memory of what still needed to be done - I put the blindfold back on and finished it just to prove it, but of course it was a DNF). For a regular 3x3x3, I still find it much faster to just do visual memorization. I keep occasionally trying to use a memory method, but I'm just to slow at it for a regular 3x3x3.
But if you want to do Roman Rooms, basically you just need to encode the information somehow and place that information at the locations in your room.
To show how I do it, for the EP example you gave, I letter 1-12 according to Macky's guide (I use Macky's method most of the time) as letters A-L. So I would start at my first location, the synthesizers, and use my person for 1 (A), "Atom Ant", and then the person for 6 (F), "Fred Flintstone". So I might have Atom Ant standing on Fred's nose while Fred plays a lick on the synthesizer. End of cycle, so I'd move on to the second location. 2 (B) = "Barney Rubble", 4 (D) and 8 (H) - I'd come up with a word to go with that 2-letter pair (that's a tough one) like maybe "dahlia", 5 (E) and 7 (G) = "egg", so I'd imagine Barney Rubble on the toy horse (my second location), holding a flower and eating an egg. Then the third location (bookcase) - 3 (C) = "Casper the friendly ghost", 9 (I) and 11 (K) = "ick", and 10 (J) and 12 (L) = "jail". I'd probably imagine Casper the ghost saying "ick" because he's been put in jail, and I'd imagine him in one of the bookcase shelves with bars in front of him, holding onto them with his tongue hanging out (saying "ick" because he's stuck there).
For just a regular 3x3x3, the cycles are all fairly short (no more than 12 for EP, less for corners), so I would pretty much never use more than one location for a single cycle. But you could, if you found it easier. So if you had an 11-cycle, you could put the first 5 at one location and the next 6 at the second location. If you're going to split a cycle like this (you pretty much need to for bigger cubes), it's helpful to use the same starting letter for each piece. So if I had an 11-cycle starting at 1 and I were going to split it, I'd have Atom Ant interact with 2 "words" in the first location, and then with 3 more "words" in the second location.
Notice how nice it is to use 2-letter pairs for a 3-cycle method like Macky's. Each "word" contains the next 2 pieces that go with your first one to make up the next 3-cycle. So you just progress through the memorization one word at a time, performing one 3-cycle per word.
To memorize a whole 3x3x3, I put EP in however many locations are needed, one per cycle, then CP in however many locations are needed, one per cycle, then use one location for EO and one more for CO, finding some "visual" way to store my hexadecimal representation of the orientations in each of those locations.
That's how I do it, anyway. There are lots of other ways you could come up with to do it - you just need to find an encoding method that works for you.
. If all else fails I can memorize where the piece goes and by knowing which sticker was first in the buffer I can find out which is now in the buffer. That requires more thinking though so hopefully it doesn't come to that.
