Some argued with me about privacy on online services. At the ideal screnario, actually online recognition is required only for LVSCR, while FSG can be handled offline if architected correctly. I think letting users to choose or not to let us use his voice to improve models is how other OSes handle this issue.
===Offline and online===
The same speech server can be designed to run both online as offline, letting the responsibility to handle transmission to the middleware that handle the connections with the front.