We develop a framework to address the shortcomings of current smartphone-based approaches for road roughness sensing and monitoring through combining vehicle dynamics, random vibration theory and a two-layer inverse analysis. The proposed approach uses in-cabin recordings of the vehicle’s vertical acceleration measured by a smartphone positioned inside the car for the estimation of road roughness. The mechanistic road roughness–vehicle interaction model at the core of the proposed framework links the frequency spectrum of the vehicle’s vertical acceleration to the road roughness power spectral density and lends itself to the quantitative characterization of roughness-induced energy dissipation. We demonstrate that the measure of roughness provided by the stochastic model of car dynamics interacting with a rough road is fully compatible, in a statistical sense, with the spatial but deterministic definition of road roughness, and validate the identification strategy that originates from it against laser measurements of road roughness. The critical crowdsourcing features of the proposed framework, such as the marginal impact of phone position and transferability, are examined and its utility to meld with big data analytics to identify the class of vehicles travelling on a roadway network is demonstrated.