This is just a guess on how you would do it... but it would involve making many many many requests until you could establish a mean against a given confidence level. Then you could compare that mean for each starting character (once again, many many many requests for each character). If the mean of any of them (or heck, even the distribution of the timings) differ than the others, then you know you you hit the right mark.

This assumes that 1: you can reliably measure at that small of a time difference and 2: you can submit enough requests without being detected.