An artificial neural network (ANN) algorithm has been applied to the automatic picking of local and regional S phase. For a set of local three-component seismic data, a variety of features for signal detection and phase identification were analyzed in terms of sensitivity and efficiency. Comparing the performance of each feature in discriminating the local S phases, four features were selected as input attributes of the ANNS-phase picker: (1) the ratio between short-term average and long-term average, (2) the ratio between horizontal power and total power, (3) autoregressive model coefficients, and (4) the short-axis incidence angle of polarization ellipsoid. The four attributes were calculated in the frequency band of 2 to 8 Hz with a 2.56-sec moving window. This choice of frequency band and window length is appropriate for local microearthquake monitoring. The results of preliminary training and testing with a set of local earthquake recordings show that the ANNS-phase picker can achieve a good performance in identification and onset-time estimation for local S phases. In overall result, 86% correct rate of phase identification has been achieved by the trained ANNS-phase picker, 74% of them are precisely picked with less than 0.10-sec onset time error. We believe that the method presented here is a promising approach to automatic phase identification and onset-time estimation.