Abstract: Accumulated observations of global earthquakes in the past decades have shown a great amount of variability of rupture behaviors; the extent that earthquakes share common rupture characteristics remains unclear. By applying a variational autoencoder to global source time function (STF) databases, we extract a generic spectrum of earthquake STF shapes along with their population density. The spectrum reveals the popularity of events with weak rupture onset and rapid termination, which contests a long-lingering perception that isosceles triangular and long-tailed shapes represent standard rupture models. The spectrum also highlights a special class of complex earthquakes that are located at shallow depth (<40 km) but pervasive in different faulting environments, implying apparent depth control on fault properties. Finally, we show that deviation from the spectrum serves as a convenient detector for earthquakes with unusual characteristics. Our results present a panoramic view of temporal rupture characteristics for global earthquakes and establish a quantitative framework to evaluate the uniqueness of individual events in the context of historical observations.