A fast reverberation suppression method based on graph Laplacian regularization

Traditional low-rank and sparse decomposition methods impose a global low-rank constraint for estimating the reverberation subspace, but they face performance degradation with limited samples and require computationally expensive singular value decomposition (SVD). This letter proposes a fast method that constructs a neighbor graph to model inter-frame relationships and introduces graph Laplacian regularization to penalize differences in reverberation between adjacent frames, thereby preserving local smoothness. The method accurately estimates the reverberation with limited samples while avoiding SVD. Experimental results show that the proposed method enhances reverberation suppression performance under limited data conditions while reducing computational time compared to existing methods.