Online speech source separation based on maximum likelihood of local Gaussian modeling

We propose an online speech source separation method which can separate sources under underdetemined conditions. The proposed method is based on local Gaussian modeling (LGM). At first, we de rive an extended approach of conventional offline speech source separation methods based on LGM, which can separate speech sources in an online manner. The likelihood function of the online LGM based approach (OLGM) is approximately maximized by incremental EM based approach. Additionally, we propose an initialization method of OLGM based on a least squares approach to improve con vergence time . Experimental results show that the proposed method can separate sources effectively even when the number of iterations is small.