A double-step grid-free method for sound source identification using deep learning

Abstract

Deep learning is a machine learning method based on the deep neural network (DNN) which is widely used in various application fields. Compared with the model-based beamforming method, the sound source identification method using deep learning is very promising. The grid-free method is one of these deep learning methods in high accuray. However, it needs to predefine a certain number of sources in advance of localization and quantization. To solve the limitation of existing grid-free methods, a double-step grid-free (DSGF) method to identify unknown number of sound sources is proposed. The classification DNN to identify sources number is trained as the first-step and 6 regression DNNs to localize and quantify sound sources are trained as the second-step. In this work, the residual neural network (ResNet) is utilized as the prediction model. Conventional beamforming (CB) map is used as input, and the output is changed with different tasks. The classification accuracy on two kinds of datasets are analyzed, then the performance of localization and quantization with correct or wrong sources number are evaluated. The results show that the DNN models trained of different tasks are reliable. Besides, the comparison beween DSGF and CLEAN-SC methods reveals that the proposed method performs better in the close-range condition. The simulation and experimental results verify the feasibility of the DSGF method.