Temporally correlated source separation using variational Bayesian learning approach

Basic blind source separation (BSS) algorithms did not adopt time information of signals. They assumed that each source was independent and identically distributed (i.i.d.). In the paper, we propose to use time structure and prior information of sources in order to improve separation. Modeling source by generalized autoregressive (GAR) process, we can tackle the problem of temporally correlated source separation using variational Bayesian (VB) learning approach. The advantages of our proposed algorithm are that (i) it makes full use of time structure of sources; (ii) it can separate different type of sources in noisy environment; (iii) it can avoid overfitting in separation. Experimental results demonstrate that our algorithm outperforms VB separation algorithm based on i.i.d. source model and second-order statistical decorrelation algorithm.