"But the more powerful systems get, the less usable VR seems to get. They also seem to get more stupid in many cases."
I think that's primarily because they've switched from recognising a few specific commands to attempting to recognise entire "natural language" sentences with little to no word gaps. It's a massive leap in complexity but is being treated like it's just a step upgrade. My Garmin SatNav recognises pretty much all of the spoken commands I give it apart from the word "up" when navigating a list. (A US style nasel "app" sound seems to work about 75% of the time) but often has some difficulty if I try to set a destination by speaking the address. It works a little better if I speak with distinct gaps between the words but does seem to have a few americanisms left in it. For examples, here in the UK we would normally give an address as One Eight Six Acacia Avenue, Anytown but the Garmin seems to prefer One Hundred Eighty SIx, Acacia Avenue, Anytown. Note the missing "and" between 100 and 80 that a Brit would normally use.
Likewise, the text to speech seems to use the US style of "I don't know that word so follow these pronunciation rules for a best guess" rather than the UK rules you'd expect of a product marketed in the UK.
I've not heard it say "wor-sester-shire" yet but Campbell comes out as Camp-Bell.