I think they've been at it for a pretty long time with old products like Dragon. A friend used to work there some years back and said they pretty much perfected speech detection up to some very reasonable error rate. I imagine they've just continued to cover all the dark corner cases and irregularities and accents etc.