Doable, but voice recognition is severely lacking outside of targeted small dictionaries, and a simple wav file can beat it. I wouldn't rely on it for a long while.

I meant like a varying question - "which is not an animal - zebra, teapot, monkey, .." so yeah, if your AI can pick the right wav file it's beatable..

In that case, yes, it would be interesting to make a system to verify such and probably doable in the short term (ignoring the UI issues which are separate). The problem still remains the same that your speech recognition is going to be very spotty on untrained processing, so you can at best (key word there) match maybe 500 words on one accent pool. This means you have a limited set of questions and it does make it a barrier to entry, but certainly not unbeatable and definitely much more in favor of the AI programmer.

This is also why your average big-company automated attendant is so bad at interpreting what you want outside of the 5 common things people want to talk about (most of which I can do online, so it's never what I'm there for) and again sadly only if you're in the more common US American English accent pool. Southern accents tend to get squelched pretty bad and that's just the beginning of accents common in the US.

One could conceivably target an individuals desired accent recognition so you could let the user say "Most of my callers speak in an X accent" and then get a little more accurate for that pool of callers, but this also means installing disparate systems for recognition, each of which are quite expensive computationally.

This is a deeply interesting problem space, though, and if you're interested there's quite a bit of discussion over it in the sphinx project and several research communities.