A neural network system for P and S-picking and location of earthquakes in Northeastern Italy is described. It is applied to 7108 seismograms corresponding to 1147 earthquakes occurring in Northeastern Italy and surrounding area in the period 2000-2003. Its results are compared with two sets of manual picks and with the picks performed by the existing seismic alert system. The new system recognizes 89% and 67% of P and S arrival times, respectively, which allows locating 92% of the earthquakes. P and S-picks differ from the best available manual picks by 0.00 ±0.07 s0.00±0.18 s, respectively. The corresponding earthquake locations differ by -0.18±0.77 km in longitude, 0.10±0.62 km in latitude and 0.1±2.0 km in depth. These results suggest its use for alert purposes and rapid production of preliminary bulletins. Considering a subset of picks that are common to all the available data sets, the absolute accuracy (i.e., the inverse of the standard deviation of differences between the estimated and the true, unknown arrival times) of each picking method is estimated. The best available manual data set has standard deviation 0.03 s for P waves and 0.07 s for S waves, while for the new system it is 0.06 s and 0.18 s for P and S waves, respectively.