Joint Estimation of Sound Source Position and Room Boundaries Using a Multitask Deep Neural Network Model

Conventional room geometry blind inference techniques with acoustic signals often rely on prior knowledge, such as source signals or source positions, limiting their applicability when the sound source is unknown. To solve this problem, the authors propose a novel multitask deep neural network (DNN) model that jointly estimates sound source localization and room geometry using signals captured by a spherical microphone array. Considering the coupling between sound source content and environmental parameters in reverberation signals, extracted early reflection direction and delay information as network inputs to estimate spatial parameters is used, ensuring independence from the sound source signal. The proposed model employs a hierarchical architecture with dedicated subnetworks to process direction-of-arrival (DOA) and time-difference-of-arrival features, followed by a shared fusion module that exploits geometric constraints between source and boundary positions. Compared with traditional methods, this model requires less prior environmental information and performs sound source localization and room geometry inference with single-position sound field measurements. Experimental results from simulations and real measurements demonstrate the method’s effectiveness and precision compared with conventional approaches across various scenarios.