The prescription of turbulent inflow boundary condition for large-eddy simulation (LES) is critical to obtain accurate predictions. In this paper, a stochastic model for generating inflow turbulence is implemented and compared to the method of generating inflow turbulence with random fluctuations. The LES studies of the effect of inflow boundary conditions are performed using a semi-implicit Legendre spectral element method. Two different backward-facing step configurations are simulated at Reynolds numbers 5000 and 28,000 utilizing the technique of implicit LES with no sub-grid model. Frequency analysis is performed at various points in the boundary layer and in the shear layer to study the effect of inflow boundary condition. The spectral method requires correlated velocity input data generated by the stochastic model for predicting the correct transition to turbulence at Re = 5000, while the random fluctuations predict a delayed transition to turbulence. At Re = 28,000, the two techniques of generating inflow turbulence give similar results. In addition, it is also observed that at Re = 5000, the semi-implicit method requires small time step size to correctly predict transition to turbulence. The large time step size leads to laminarization of the flow in the inlet channel. This effect of the time step size is not observed in the case of Re = 28,000. Published by Elsevier Ltd.