[4N3-GS-7] Vision, speech media processing:

[4N3-GS-7-04] Reduction of Measurement Points in Sound Field Reproduction Using CNN with Skip Connections

〇Koki Horikoshi¹, Gen Sato¹, Izumi Tsunokuni¹, Yusuke Ikeda¹ (1. Tokyo Denki University)

Keywords:U-Net, ResNet, Pressure Matching

Sound field reproduction is a technique used to reproduce acoustic spaces using multiple loudspeakers. The Pressure Matching (PM) method, one of the sound field reproduction techniques, enables high-accuracy sound field. However, obtaining the transfer functions of the loudspeaker and the desired sound pressures requires many measurement points, making practical implementation challenging. Therefore, in previous researches, the number of measurement points was reduced while maintaining high accuracy through a data-driven deep learning PM (DLPM) method using convolutional neural network (CNN). In this study, we propose a DLPM method that applies Res-UNet, which integrates ResNet and U-Net, to improve accuracy with fewer measurement points. We conducted simulation experiments for evaluation. In the experiments, we compared the performance of different network architectures including CNN, ResNet, U-Net, and Res-UNet. Deep networks are constructed using residual connections, while skip connections between distant encoder and decoder layers help prevent gradient vanishing.