Blind Source Separation of Speech and Acoustic Signal for Moving Audio Sources

摘要

We execute ICA of different type to an artificial mixed audiornsignals and also to audio signals in real environment.rnThe time-domain ICA uses the second order cross-correlationrnwith non-stationary assumption. For the frequency-domain ICA,rnwe use infomax criterion generalized to the complex domainrnand Kullback-Leibler information.rnFor our artificial data from [6], every our method makesrnsatisfactory separation as preliminary trials.rnOn the other hand, our experiments in the real environment dornnot show enough separation between two original sources, inrnorder to seek the reason for which is our next problem.