So upon reading back, it appears that you have eliminated the backboard, and now the I/O board, so yeah, I am with you believing it is the MPU. I do not have a breakdown of what exactly on that board is dedicated to sound, but obviously the XU17, XU18, XU30, and XU31 eproms are all sound chips. With that knowledge, you can follow the traces and see what route they take to various other chips. A bit of work, but at least there is a visual map.
Again, with the knowledge that the sound is not working well when cold and then starts working better when the machine has generated some heat, my assumption is that there is or are cold solder joints on the MPU somewhere. Look for those on the back and see if you can locate any, maybe at the feet of the sound eprom sockets, or wherever the traces may travel from there. When the machine cools, the solder shrinks enough to lose some contact somewhere along the way and when it heats, it expands to complete the contact. This is a common problem in older pinball machines as well that we run into regularly. It's the only thing I can think of at this point. I would start with reflowing the solder for the eprom sockets located at 17, 18, 30, and 31 and then go from there. It is definitely not an eprom programming issue, because it would either work or not work. There is no way that would cause intermittent problems.