Blind Source Separation of Audios Based on Bayesian Method

Noise interferences and the lack of observation data are two main problems when implementing blind source separation. This paper introduces a BASS method which eliminates noise interferences, and meanwhile, uses sparse component analysis to fit the undetermined situation. Firstly, based on the BASS mathematical model including linear noise affixed, we built a probability distributed model of the sparse representation parameters of original signals, which is obtained by Wavelet transformation of original signals; then, we adopt the Gibbs sampling algorithm to estimate the parameters by alternate calculations for each parameter, and finally we get the separation result in an iterative way. Experiment results show the advantage of our algorithm for percussion music compared with other traditional methods especially in undetermined blind separation problems.