In the numerical computation of wave equations, numerical dispersion is a persistent problem arising from inadequate discretization of the continuous wave equation. To thoroughly understand the mechanism of numerical dispersion, we separately analyze the numerical dispersion relations of time-stepping and spatial discretization schemes by Fourier analysis. The relevant results show that the numerical dispersion errors of time-marching schemes depend on the time step length or the Courant number, whereas the numerical dispersion errors of spatial discretization schemes are determined by the error between the eigenvalues of the numerical spatial differential operator and the continuous spatial differential operator. We also find that the much better numerical dispersion accuracy of the stereo-modeling discrete (SMD)-type operator can be attributed to the inclusion of diversified basis functions for its eigenvalue. Based on these findings, we combine the optimal four-stage symplectic partitioned Runge-Kutta and eighth-order SMD as a new fully discrete scheme. The subsequent analysis of its normalized phase velocity is consistent with the numerical dispersion analysis in semidiscrete forms. This is followed by an acoustic wave simulation in a homogeneous model and corresponding computational efficiency comparison. The results show that the new scheme is much more accurate and antidispersive on a coarse grid. In the final two numerical experiments, we use the new scheme to model the acoustic wave propagation in a three-layer model and the Marmousi model. The convolutional perfectly matched layer is applied to eliminate artificial boundary reflections. Our semidiscrete numerical dispersion analysis provides an efficient tool to quantitatively evaluate the time-stepping and spatial discretization schemes. It can facilitate the development of more accurate fully discrete numerical schemes for solving seismic wave equations.