Matrix Estimation Based on Single Source Labeling

Abstract

In order to solve the problem of matrix estimation in underdetermined blind source separation, a matrix estimation method based on single source labeling was proposed. Firstly, the unknown mixed signals of each channel were predicted and classified by semantic segmentation network, and the single source points and mixed time-frequency points were labeled respectively. The threshold was set to obtain the high-score single source points of each category. Then, the interval probability statistical detection method was used to remove the interference points in semantic single source points, improve the accuracy of mixed matrix estimation. The experimental results show that the accuracy of the proposed algorithm in estimating the number of source signals at low SNR is much higher than that of the comparison algorithm, and the normalized mean square error of the estimation matrix is about 4 dB lower than that of the comparison algorithm.