The large amount of digital data recorded by permanent and temporary seismic networks makes automatic analysis of seismograms and automatic wave onset time picking schemes of great importance for timely and accurate event locations. We propose a fast and efficient P- and S-wave onset time, automatic detection method based on neural networks. The neural networks adopted here are particular neural trees, called IUANT2, characterized by a high generalization capability. Comparison between neural network automatic onset picking and standard, manual methods, shows that the technique presented here is generally robust and that it is capable to correctly identify phase-types while providing estimates of their accuracies. In addition, the automatic post processing method applied here can remove the ambiguity deriving from the incorrect association of events occurring closely in time. We have tested the methodology against standard STA/LTA phase picks and found that this neural approach performs better especially for low signal-to-noise ratios. We adopt the recall, precision and accuracy estimators to appraise objectively the results and compare them with those obtained with other methodologies. Tests of the proposed method are presented for 342 earthquakes recorded by 23 different stations (about 5000 traces). Our results show that the distribution of the differences between manual and automatic picking has a standard deviation of 0.064 s and 0.11 s for the P and the S waves, respectively. Our results show also that the number of false alarms deriving from incorrect detection is small and, thus, that the method is inherently robust.