Aerodynamic noise generation in a high-speed train bogie compartment is closely associated with the interaction between the shear layer developing from the cavity leading edge and the wakes generated by the wheelset, brake disks, and bogie frame. This study introduces square-tooth leading-edge spoilers (STLESs) with inclination angles of β = 30°, 60°, and 90° as controlled geometric perturbations to investigate how spanwise modulation alters coherent flow structures, unsteady surface loading, equivalent surface-dipole source-power characteristics, and far-field noise radiation at 400 km/h. Large-eddy simulation is coupled with the permeable Ffowcs Williams–Hawkings analogy, equivalent surface-dipole source-power analysis, and spectral proper orthogonal decomposition (SPOD). The spoiler-free prototype model exhibits pronounced spanwise-coherent pressure fluctuations at the fundamental frequency f0 = 547 Hz and its harmonic 2f0 = 1094 Hz. STLES-30 disrupts this coherent loading, fragments elongated vortical structures within the wheelset–frame interaction region, and reduces the energy contribution of the leading velocity SPOD mode from 35.4% to 28.0%. At 547 Hz, the leading pressure-SPOD eigenvalues associated with the brake disk, axle, and wheel sections decrease by 22.6%, 15.6%, and 3.8%, respectively; at 1094 Hz, the leading eigenvalue decreases by 40.6%, accompanied by a redistribution of modal energy toward higher-order modes. These modifications to the coherent flow and pressure fields weaken the dominant equivalent surface-dipole source contributions at 250, 500, and 1000 Hz and reduce the overall sound pressure level at all 16 standard receivers, with a maximum reduction of 2.2 dB. Larger square-tooth angles induce stronger spanwise modulation but may also introduce additional pronounced peaks in the equivalent source-power spectra. The results indicate that effective passive noise control depends on sufficiently disrupting coherent shear-layer loading while avoiding the formation of new localized strong equivalent source regions.