In seismic tomography, traveltime infor mation of seismic body phases is commonly used to invert the seismic velocities of the subsurface structure. At long periods or for later seismic phases, the arri v al time of seismic phases lack definiti ve onset and a direct picking of the absolute arri v al time has large uncertainty and reproducibility. A common practice is to estimate the relative delay between the observed and synthetic signals that maximizes the correlation coefficient. For that aim, we must first select appropriate time windows around the candidate signals. To improve the ability to detect and extract weak signals, we develop a new morphological time window selection (MTWS) algorithm that adapts to the shape of signals and has robust performance in automated processing of massive data. The MTWS method consists of two successive steps. First, we detect the major peaks on the waveform envelope using a maximum filter. Secondly, we solve for the beginning and end of the time windows surrounding the peaks straightforw ardl y from simple geometrical equations. The efficiency and robustness of the MTWS algorithm make it very suitable for automated processing of huge data sets. We demonstrate the implementation of the method with both synthetic and observed long period (20-40 s) SH waves. From ∼100 000 traces of transverse-component seismograms recorded by global seismic networks over the course of a year, we obtain ∼15 000 S diff , ∼7500 ScS and also some ScS multiples. The global map of S diff correlation time delays shows consistent patterns with the shear wave velocity perturbations on the core-mantle boundary in the recent tomographic models.