The structural design of gearbox housing, such as ribbing and wall thickness, has a significant impact on its noise radiation characteristics, especially in electric vehicle applications where tonal noise is more perceptible. This study presents a novel methodology that uses machine learning and spectral analysis to distinguish between gearbox housing types based solely on their acoustic radiation data. Frequency-domain sound pressure spectra, simulated for multiple design variants, were interpolated and analyzed using Principal Component Analysis (PCA) and K-means clustering. The results reveal that construction types (e.g., fully ribbed, partially ribbed, or without ribs) exhibit distinct acoustic profiles. Furthermore, a Random Forest classifier achieved 88.9% accuracy in predicting structural configuration from the spectra alone. These findings demonstrate that structural design features can be inferred directly from acoustic data, offering a lightweight and geometry-free alternative to traditional NVH simulation workflows. The approach can be integrated as a lightweight plug-in in existing NVH workflows. It ingests acoustic spectra and returns a structural-stiffness label with uncertainty, supporting early-stage screening and late-phase regression checks.