NMF2D-based source separation using extreme learning machine

In this paper, we study Non-negative Matrix Two-Dimensional Factorization (NMF2D) based Single Channel Source Separation (SCSS) using a newly proposed algorithm named Extreme Learning Machine (ELM). Compared with other machine learning algorithms such as Support Vector Machines and Neural Networks, ELM can provide better generalization performance and a much faster learning speed. Unlike conventional researches that concentrate on generating masks for each source, we use ELM to classify estimated sources separated by NMF2D algorithm. We also explore Deep ELM which means more than one hidden layers to improve the performance. While training Deep ELM, a method named layer by layer pre-Training is used, but unlike Deep Belief Networks (DNNs) that need to fine-Tune the whole network at the end, Deep ELM can be used without iteration fine-Tuning. The experiment results show that the performance of proposed method is improved not only in training and testing speed, but also in the quality of separated signal compared with using DNNs and NMF2D.