The short answer is that accounting for those sorts of differences is really damned hard and represents most of the research effort in the field for the last forty years.
The long answer is that modern voice recognition works off of a giant probability graph that matches input to the most likely sequence of phonemes (and the most likely sequence of words, from there). If you're trying to match against the entire vocabulary of a language your error rates are still fairly substantial, but cutting the vocabulary way down like you would for a device like Kinect results in many less possibilities, which lets you achieve much better accuracy even across multiple speakers.
A dialect so thick that even people have trouble with it isn't likely to play well with a device expecting English input, however.
Read up on hidden Markov models and how they're applied to speech recognition if you want to get more into the science behind the tech.
I will. It's not something we talked about back when I was in undergrad, and I was graduating as they were about to release Kinect, so I was intrigued back then. Since I switched fields of study, I haven't been able to keep up with it. Sound studies (ex: optimality theory -- so the study of sound rules and phonotactics), historical linguistics, and neuro/psycholinguistics, yes, but other fields that I wasn't too experienced on, no.
But yes, accounting for all of those differences is incredibly difficult to do. Especially when you take into account all of the languages in existence worldwide, and
especially when you take into account accents, dialects, pidgins, creoles, etc. It's just hard to accommodate for every single speaker in the world. There are even individual-to-individual variances, so the tech isn't always going to be perfect. Using common words like they usually do in the conferences certainly work. "Xbox play! / Xbox joué! / Xbox juega!" or something like that would work far more based on the lack of tons of variables at work. I would expect that they're using a
standard with respect to that, so the most out-there accents or dialects mightn't work.
I do remember that they only had certain dialects of Spanish present, for example. So Spain Spanish as opposed to other forms of Spanish. And sometimes it didn't work since the phonological input was not what Kinect was expecting.
Anyway, I just find that aspect of the tech interesting.
