I wonder if NLP research should have started as our ancestors did, with grunts and hoots and cries. Instead it's focused on recognizing full words and sentences while almost completely ignoring inflection.
Another dimension to add with vocal input is directional. If you have mics in all corners of a room, which direction you speak in can affect whether "turn off" operates your TV, your lights or your oven.
There is a lot of improvement for voice processing in several dimensions of voice.
Taking sections of the last response, or hell, even having every response essentially be wrapped up in some sort of object you can reference in your next query to the interface is what all of these lack.
Even Androids "Search this artist" doesn't quite get there. The lack of context between queries is what murders Siri for me. That and her seemingly random selection of what goes to Google and what goes to Wolphram Alpha. Sometimes even the "wolfram" verb prepended to a query just doesn't go to wolfram no matter what.
(I'd suggest this is actually a combination of the still-non-trivial nature of NLP, combined with a lack of feedback, combined with the fact that giving instructions is quite hard. Humans overestimate human language's ability to communicate clear directions, as anyone who has done tech support over a phone understands.)