An online adaptive algorithm for underdetermined blind source separation

Underdetermined blind source separation (UBSS) deals with the problem of estimating n source signals from m measurements (n > m), with an unknown mixing process. Most researches pay attention to the sparsity of speech to recover source signals, such as the DUET (degenerate unmixing estimation technique) algorithm, which can separate any number of sources using only two mixtures with the help of estimation of mixing parameters but cant be used in the time-varying system. We develop a block-online algorithm for a dynamic environment that is based on mixing parameter estimation from reliable time-frequency points determined by the total power and FCM (Fuzzy C-Means) clustering of parameter estimates for every time-frame. To track the variable mixing parameters, the estimated ones from the previous frame can be used to initialize FCM in the current frame with changed mixing measure instead of random initialization in case of the conventional FCM. Thus, an online adaptive algorithm of UBSS with the time-varying matrix is realized.