Acoustic source localization for a single point source using convolutional neural network and weighted frequency loss

초록

In this paper, a deep learning-based acoustic source localization model is proposed to estimate the location and strength of the sound source accurately. Conventional model-based inverse approaches, e.g., beamforming and deconvolution, have limits in terms of accuracy, computational cost, and dependency from domain knowledge. Recently, deep learning-based methods have been applied for the acoustic source localization to reach localization performance better than conventional methods. Although several pieces of research have been suggested to achieve the deep learning-based sound source localization, there is still a performance-related issue based on frequency for the deep learning model. In order to overcome the limitation of the previous works, we mainly utilize a convolutional neural network (CNN) designed to conduct both feature extraction and feature learning of spatial information based on the beamforming maps from low to high frequency. Besides, a very deep CNN model with dense connectivity among feature maps is established to maximize channel-wise residual mechanism, as well as encourage effective feature reuse and feature propagation. In particular, we suggest the weighted frequency loss for training the deep neural network by incorporating frequency information of the sound source into the loss function. We observe that our proposed approach generally enhances the model's prediction performance according to the frequency and also reduces the difference in localization performance between low and high frequencies. © Proceedings of 2020 International Congress on Noise Control Engineering, INTER-NOISE 2020. All rights reserved.