An improved cosine single source point detection algorithm for mixing matrix estimation

Precise estimation of the mixing matrix is essential for efficient recovery of source signals in underdetermined blind source separation (UBSS). In light of the limitations in either precision or noise robustness exhibited by existing approaches for mixing matrix estimation (MME), an improved cosine single-source point (SSP) detection algorithm is proposed in this paper. The proposed method comprises two main stages: (1) extracting initial SSPs from the mixed signals using cosine SSP detection, and (2) identifying and eliminating abnormal points in the initial SSPs by analyzing their linear relationships. After completing the SSP detection, the K-means algorithm is employed to cluster the refined SSPs in order to obtain MME. Experimental results demonstrate that the proposed algorithm significantly enhances the accuracy of SSP detection using environmental sound and speech signals. Compared to the cosine SSP detection algorithm, the normalized mean square error of the MME has improved by approximately -9.9903 dB. Further experiments show that, given equivalent signal-to-noise ratio, the proposed method outperforms existing approaches in terms of MME accuracy.