Acoustic Source Localization for Single Point Source using Convolutional Neural Network and Weighted Frequency Loss

In this paper, a deep learning-based acoustic source localization model is proposed to estimate the location and strength of the sound source accurately. Even though a variety of existing works have been proposed to achieve localization and characterization of the sound source, the conventional model-based inverse approaches, e.g., beamforming and deconvolution, have limitations in terms of accuracy, computational cost, as well as a dependency from domain knowledge. In order to overcome the existing problems, a convolutional neural network (CNN) is utilized to carry out both feature extraction and feature learning of spatial information based on the beamforming maps from low to high-frequency bands. Moreover, the CNN model with dense connectivity among feature maps is adopted to maximize channel-wise residual mechanism, as well as encourage effective feature reuse and feature propagation. In particular, we suggest the weighted frequency band loss for training the neural network by incorporating frequency band information of the sound source into the loss function and resulting in productive learning, regardless of frequency band ranges.