Comparison of the room’s dimensions and absorption distribution estimation performance using wave-based and geometrical acoustics dataset

Deep neural networks (DNNs) are trained to extract the room dimensions and absorption configurations from room transfer function (TF) measurements. This study investigates the performance of DNNs in room acoustic analyses, which are trained with wave-based (WB) and geometrical acoustics (GA) simulation data. WB simulation data provide a physically accurate representation of room acoustics including diffraction and interference, albeit with substantial computation demands. In contrast, GA data can be obtained more rapidly, but with reduced accuracy. We found that the DNN trained with WB training data exhibits enhanced estimation performance and generalization capabilities when applied to real-world measurements. This study underscores the trade-offs between training dataset generation speed and their performance of machine learning algorithms in the inverse problem