Right, the missing step to make these assistants actually-useful beyond very simple tasks like "play [album name]" or "set a timer for 5 minutes"—which are basically just a voice-driven command line—has almost nothing to do with voice recognition per se and is more-or-less identical to the problem of "invent AGI".

It'll be useful when you can have it create, say, a menu and shopping list for your week, interact conversationally to tweak it including understanding that something like "I'm allergic to that ingredient" means never ever try to include it again while other statements of preference may be much softer, learn your preferences so those conversations get shorter every week, and order the ingredients for you, too. It barely matters if that happens over keyboard input or voice, because the hard parts of that problem have nothing to do with how the input is done.