Initial test showed over 95% accuracy, and we are working on make it even better. On one hand it's not intends to be instead of your GPS to count each and every run mileage, even though you can. BTW, it will clearly count also treadmill runs, while a GPS will not...

After training on three five mile data sets, I was able to get very accurate speed/distance estimates for subsequent runs (say 99%+).

The method? 1st derivative of accelerometer data to determine cadence. Then correlate cadence with actual speed with a simple linear regression when GPS data was available.

This was more intended for dead reckoning on a wrist mounted device, though. We had a 3 axis accelerometer and a 3 axis magnetometer.