Evaluation of Data-Driven Room Geometry Inference Methods Using a Smart Speaker Prototype

Recent studies tackling the problem of room geometry inference (RGI) with data-driven methods require a substantial amount of room impulse responses (RIRs) collected in a diverse set of rooms for training the deep neural networks (DNNs). However, this may be a prohibitively time-consuming and labor-intensive task, which requires simulated data. This study explores regularization methods to improve RGI accuracy when DNNs are trained with simulated data and tested with measured data. We use a smart speaker prototype equipped with multiple microphones and directional loudspeakers for real-world RIR measurements. The results indicate that applying dropout at the network's input layer results in improved generalization compared to using it solely in the hidden layers. Moreover, RGI using multiple directional loudspeakers leads to increased estimation accuracy when compared to the single loudspeaker case, mitigating the impact of source directivity.