Have had same concerns as far as too much voice info throughout the house. I have a panasonic system, hai and doorbell output module set up through paging in the cav66 system. 18 zones now, getting ready to add 6 more. Originally I was going to add the mlspeak to the whole house paging mode but quickly realized, WAF NOT!!, that it would not work. I now am playing with putting the mlspeak as one of the inputs into the cav66 and using the cav plug in to direct where the sound goes. I can use voice response prompts in just the zone where the computer running the ml scene is located. I can direct certain responses to just certain zones. Missed calls to the garage upon arrival, etc. Wake up alarms to just the zones awakening. I can control, depending on time and other factors, which zones get full house pages. Possibilities are really endless depending on how much programming you want to do. Using recursive loops in command maps ( indebted to bhiga for his tutorial on his website ) it is fairly easy to record the on-off status, input selection and volume for each zone, select input MLSPEAK, say the message and then return the zones to their previous state. One command repeated for each zone. Easy to update and change.
This indirectly solved another problem I was having. If I announced callerID overhead, one statement was not enough, so I repeated it three times. This would then be distracting to pick up the call because it would continue to talk overhead. Since it was thru the page input it could not be muted with a phone off hook variable. Now since it is an input selection the callerid is announced until the phone is picked up and then muted, as is the music in general.
With a Panasonic interface then just the zone where the phone is picked up could be muted.
With an RFID interface then the music zone could follow where the wireless tablet goes.
Still testing and a long way away from any kind of finished product but very WAF accepted and blows my friends AMX and crestron systems away.