A temporary seismic array was in operation between October 2010 and March 2013 in the Gyeongju area of Korea. Teleseismic records of the seismic array appropriate for receiver function analysis were collected, and selected seismograms were split into five groups based on epicenters—the Banda-Molucca, Sumatra, Iran, Aleutian, and Vanuatu groups. 1D velocity structures beneath each seismic station were estimated by inverting the stacked receiver functions for possible groups. The inversion was done by applying a genetic algorithm, whereas surface wave dispersion data were used as constraints to avoid non-uniqueness in the inversion. The composite velocity structure was constructed by averaging the velocity structures weighted by the number of receiver functions used in stacking. The uncertainty analysis for the velocity structures showed that the average of 95% confidence intervals was ±0.1km/s. The 3D velocity structure was modeled through interpolation of 1D composite velocity structures. Moho depths were determined in each composite velocity structure based on the AK135-F S-wave velocity model, and the depths were similar to the H-κ analysis results. The deepest Moho depth in the study area was found to be 31.9km, and the shallowest, was 25.9km. The Moho discontinuity dips in a southwestward direction beneath the area. A low velocity layer was also detected between 4 and 14km depth. Adakitic intrusions and/or a high geothermal gradient appear to be the causes of this low velocity layer. The 3D velocity structure can be used to reliably assess seismic hazards in this area.