I guess if you have a central SR processor which everyone's connecting with (as per proposed solution for issue 1) then you also have a huge amount of incoming

I guess if you have a central SR processor which everyone's connecting with (as per proposed solution for issue 1) then you also have a huge amount of incoming data from all of your users. That means your system can evolve by taking feedback on errors its made, and can rapidly improve based on this new information, building larger dictionaries, taking info such as location data to improve context awareness, and using the vast volume of samples to improve SR where unusual accents are involved.