We use the new global database of source time functions (STFs) and focal mechanisms proposed by Vallée (2013) using the automatic SCARDEC method (Vallée et al., 2011) to constrain earthquake rupture duration and variability. This database has the advantage of being very consistent since all the events with moment magnitudes Mw>5.8 that have occurred during the last 20 years were reanalyzed with the same method and the same station configuration. We analyze 1754 shallow earthquakes (depth<35 km) and use high‐quality criteria for the STFs, which result in the selection of 660 events. Among these, 313 occurred on the subduction interface (SUB events) and 347 outside (NOT‐SUB events). We obtain that for a given magnitude, STF duration is log normally distributed and that STFs are longer for SUB than NOT‐SUB events. We then estimate the stress drop using a proxy for the rupture process duration obtained from the measurement of the maximum amplitude of the STF. The resulting stress drop is independent of magnitude and is about 2.5 times smaller for the subduction events compared with the other events. Assuming a constant rupture velocity and source model, the resulting standard deviation of the stress drop is 1.13 for the total dataset (natural log), and about 1 for separate datasets. These values are significantly lower than the ones generally obtained from corner‐frequency analyses with global databases (∼1.5 for Allmann and Shearer, 2009) and are closer to the values inferred from strong‐motion measurements (∼0.5 as reported by Cotton et al., 2013). This indicates that the epistemic variability is reduced by the use of STF properties, which allows us to better approach the natural variability of the source process, related to stress‐drop variability and/or variation in the rupture velocity.