Reverberation aware deep learning for environment tolerant microphone array DOA estimation

Abstract

By transferring the direction of arrival (DOA) estimation into a problem of spatial sample classification, deep neural network (DNN) driven by large training data have been investigated to achieve significant performance enhancement compared to classic DOA estimation strategies do. However, consider the highly diverse indoor environments that microphone arrays being used, different reverberation pattern under various environment pose extremely significant limitations to the DOA network model, as its performance is dependent on the match between training data and practical data. In order to improve the robustness of DOA estimation, a reverberation aware network (RAN) algorithm is proposed in this paper. The beam cross-correlation (BCC) is defined as the cross-correlation between the beamforming output and the observed multi-channel signal, which carries an approximation of the room impulse response. Adopting covariance matrix of BCC matrix as the DNN input to enable explicit characterization of reverberation in which the speech was captured, a reverberation aware deep learning network is trained to improve the generalization capability. Finally, simulation and experiment are performed to verify the effectiveness of the proposed method in improving the environment tolerance of DOA estimation, by comparing to the classic traditional DOA estimation algorithm and MLP strategy.