>>11
Quite significantly on any input that is hard enough to read that there is a chance of getting it wrong. There's a reason numeric forms are routinely read with machines now, while text OCR is still mostly limited to stuff like addresses with their limited domain, and large-scale search where errors do not matter much.