Detecting deep voids in urban subsurfaces remains a challenging problem for geophysical methods, particularly with conventional active seismic sources that lack the low-frequency energy necessary for greater depth penetration. To address this, we explored an alternative approach using seismic noise generated by the engine of an idle vehicle to supply the needed low frequencies and enhance the detection capabilities for voids at depths beyond 8 m. A key obstacle in processing this seismic noise is the presence of cross-correlation residuals, which occur when traditional linear stacking methods struggle to achieve stable averages, particularly with shorter-than-usual time duration data. These residuals interfere with the identification of void signatures, such as backscatters and ringing patterns, undermining the accuracy of void detection in dynamic urban settings. To tackle this issue, we employ symmetric autoencoders (SymAE), an unsupervised machine learning architecture, applied to vehicle seismic noise. By training neural networks to disentangle coherent subsurface information from source-induced noise, SymAE effectively minimizes cross-correlation residuals and enhances the extraction of backscatters and ringing patterns. By applying this technique in a semi-controlled field experiment, we successfully detected two known voids at approximately 10 m depth. SymAE reduces data requirements while improving detection accuracy, showing potential for rapid, cost-effective, and reliable subsurface void detection in urban environments.