Multi-channel blind speech separation method for diffuse noise environments

Independent vector extraction is an advanced method for blind source separation (BSS) in diffuse noise environments, suitable for overdetermined systems and characterized by high computational efficiency. However, it is limited by the assumption of stationary Gaussian noise, which makes it ineffective at eliminating diffuse noise that is aligned with the target source. To tackle this issue, a multi-channel blind source separation method is proposed for diffuse noise environments. This method assumes that the energy distribution of diffuse noise is uniform in all directions and exhibits time-varying characteristics. A low-rank source model and a rank-1 spatial model are utilized to construct a probabilistic model for the noisy mixtures. Using this probabilistic model, the update formula for the separation matrix is derived based on the maximum likelihood criterion, and the power spectral densities of the speech and noise components are estimated. Subsequently, Wiener filtering is employed to suppress the noise components that are aligned with the target source direction. Ultimately, experimental results demonstrate that the proposed method significantly outperforms existing BSS algorithms in terms of source separation performance and noise suppression capability, thus validating its effectiveness in complex acoustic environments.