Deep learning-based speech separation models are typically trained with simulated room reverberation or without reverberation, as realistic reverberation data are not generally available and are time-consuming to collect. However, the extent to which models trained with synthetic reverberation generalize to realistic reverberant conditions, and how this depends on acoustic-condition mismatch, remains underexplored. In this paper, we systematically compare models trained under anechoic, simulated, and real reverberation conditions for joint speech separation and dereverberation. Furthermore, we evaluate performance under matched and mismatched acoustic conditions by training on subsets of reverberation time (RT60) values, to assess the degree to which the acoustical properties must be aligned across training and evaluation conditions. Our results show that models trained with simulated reverberation can perform as well as models trained on real reverberation, particularly when trained on matched or higher RT60 values. In contrast, models trained under anechoic conditions consistently underperform in realistic reverberant environments. These findings suggest that using synthetic reverberation for training speech separation models is a viable and effective alternative to using realistic reverberation.