Prediction of the sound field in large urban environments has been limited thus far by the heavy computational requirements of conventional numerical methods such as boundary element (BE), finite-difference time-domain (FDTD), or ray-tracing methods. Recently, a considerable amount of work has been devoted to developing energy-based methods for this application, and results have shown the potential to compete with conventional methods. However, these developments have been limited to two-dimensional (2-D) studies (along street axes), and no real description of the phenomena at issue has been exposed (e.g., diffraction effects on the predictions). The main objectives of the present work were (i) to evaluate the feasibility of an energy-based method, the diffusion model (DM), for sound-field predictions in large, 3-D complex urban environments, (ii) to propose a numerical hybrid method that could improve the accuracy and computational time of these predictions, and (iii) to verify the proposed hybrid method against conventional numerical methods. The proposed numerical hybrid method consists of a full-wave model coupled with an energy-based model. The full-wave model is used for predicting sound propagation (i) near the source, where constructive and destructive interactions between waves are substantial, and (ii) outside the cluttered environment, where free-field-like conditions apply. The energy-based model is used in regions where diffusion conditions are met. The hybrid approach, as implemented in this work, is a combination of FDTD and DM models. Results from this work show the role played by diffraction near buildings edges close to the source and near the exterior boundaries of the computational domain, and its impact on the predictions. A wrong modeling of the diffraction effects in the environment leads to significant under or overpredictions of the sound levels in some regions, as compared to conventional numerical methods (in these regions, some differences are as high as 10 dB). The implementation of the hybrid method, verified against a full FDTD model, shows a significant improvement of the predictions. The mean error thus obtained inside the cluttered region of the environment is 1.5 dB. / Master of Science