Unlike air pollution, traffic-related noise remains unregulated and has been under-studied despite evidence of its deleterious health impacts. To characterize population exposure to traffic noise, both acoustic-based numerical models and data-driven statistical approaches can generate estimates over large urban areas. The aim of this work is to formally compare the performances of the most common traffic noise models by evaluating their estimates for different categories of roads and validating them against a unique dataset of measured noise in Long Beach, California. Specifically, a statistical land use regression model, an extreme gradient boosting machine learning model (XGB), and three numerical/acoustic traffic noise models: the US Noise Model (FHWA-TNM2.5), a commercial noise model (CadnaA), and an open-source European model (Harmonoise) were optimized and compared. The results demonstrate that XGB and CadnaA were the most effective models for estimating traffic noise, and they are particularly adept at differentiating noise levels on different categories of road.