A hybrid empirical method is proposed to simulate broadband strong ground motion that combines a kinematic complex source model with both numerical and empirical Green’s functions. The kinematic approach is based on a composite source model description where subevents are generated using a fractal distribution of sizes. The subevents are set up with a size‐dependent rise time. Each elementary source is described as a crack‐type slip model that starts radiating when the rupture front reaches the nucleation point located randomly inside a size‐dependent nucleation region. The synthetic acceleration spectra follow the ω2 model, and the spectral amplitudes are scaled by a frequency‐dependent directivity coefficient. In this study, the hypothesis of constant stress drop of subevents is released to better model the high‐frequency level radiated by the source. Taking advantage of small‐magnitude events, synthetic seismograms are computed using hybrid Green’s functions (HGF) to model the impulsive response of the medium. The procedure computes HGF for each subfault combining the numerical low‐frequency and the empirical high‐frequency Green’s functions with appropriate amplitude and phase correction (geometrical spreading and time delays due to the S‐wave travel‐time propagation). This methodology is applied to simulate the strong ground motions recorded during the 1997 Mw 5.9 Yamaguchi‐Ken, Hokubu, Japan, earthquake. Some of the finite‐source rupture parameters (e.g., rupture velocity, asperity size, and seismic moment) were chosen according to previously published studies conducted on this event. Only random generation of k−2 composite slip distributions was allowed (k being the radial wavenumber), to model heterogeneous slip while mimicking the main asperity size and location imaged. Comparisons of predicted and observed strong‐motion characteristics show that predictions are largely improved, when both HGF and variable stress drop of subevents are used.