Robust Demixing Filter Update Algorithm Based on Microphone-wise Coordinate Descent for Independent Deeply Learned Matrix Analysis

In this paper, we propose a robust demixing filter update algorithm for audio source separation, which is the task of recovering source signals from multichannel mixtures observed in a microphone array. Recently, independent deeply learned matrix analysis (IDLMA) has been proposed as a state-of-the-art separation method. IDLMA utilizes the deep neural network (DNN) inference of source models and the blind estimation of demixing filters based on sources' independence. In conventional IDLMA, iterative projection (IP) is exploited to estimate the demixing filters. Although IP is a fast algorithm, when a specific source model is not accurate owing to an unfavorable SNR condition, the subsequent update of filters will fail. This is because IP updates the demixing filters in a sourcewise manner, where only one source model is used for each update. In this paper, we derive a new microphone-wise update algorithm that exploits all information of the source models simultaneously for each update. The microphone-wise update problem cannot be solved by IP, but instead, a new type of vectorwise coordinate descent algorithm is introduced into the proposed algorithm to realize convergence-guaranteed parameter estimation. Experimental results show that the proposed update algorithm achieves better separation performance than IP.