A Deep Learning Method for DOA Estimation with Normalized Features in Reverberant Environment
Weilun Fang, Qinghua Huang, Zhengyu Chen · 2022
Deep learning framework has been utilized to estimate the Direction-of-Arrival (DOA) with spherical microphone array under environments with reverberation and noise for low computational complexity and high accuracy. This paper proposed a convolutional neural network (CNN) with normalized input features for classification. The received signal is first normalized along microphones to obtain the first normalized features. Secondly normalize the received signal along the spherical harmonic order to form the second normalized features. Finally assemble the two normalized features to get the input features to the network. The classifier outputs the posterior probabilities of all DOA candidates according to the input feature and determines the DOA position through the principle of maximum posterior probability. This method effectively provided the information of the DOA along microphones, spherical harmonic order, which is useful for improving the accuracy. Experiments are conducted to show that the proposed method has higher accuracy than the conventional methods and can effectively resist the reverberation and noise.