Noise, Vibration, and Harshness (NVH) performance is a key aspect to evaluate passengers’ comfort in vehicles. At the gearbox level, the gear tooth profile deviations give rise to vibrations during the gear meshing process, which is the source of gear whine noise. Therefore, an NVH design should account for the uncertainties at the gear micro-geometry level, demanding several evaluations of the gearbox model. Our study introduces a gearbox surrogate model based on Neural Networks (NN) to create a fast and accurate copy of the static and dynamic gearbox simulations made in Romax software. An accurate surrogate model was built, enabling a reduction of uncertainty propagation analyses time by a factor of 380, compared with traditional modeling by Romax.