Noise-Robust Estimation of Early-part Room Impulse Responses based on Physics-Informed Neural Network with Dynamic Pulling Method

Recently, physics-informed neural networks (PINN) has been applied to the problem of estimating the early room impulse responses (RIR) from a small number of microphones. PINN uses two types of loss functions corresponding to physical laws and data errors. Thus, PINN has the challenge that when gradients associated with two different loss functions conflict with each other, they are not properly learned. In this study, we propose a method for estimating early-part RIRs by introducing the dynamic pulling method (DPM) into the SIREN architecture with residual connections, using noisy microphone signals. From the simulation experiments on a two-dimensional sound field, the proposed method was able to estimate RIRs with higher accuracy in noisy environments than conventional methods.