Real-time blind separation of acoustic signals based on physical and geometrical properties of the sources of the signals

Blind signal separation is a technical problem of identifying separate sources from an acoustic mixture of signals. Efficient signal separation allows for development of various sensor systems to be used in clinical, industrial or home environment. It is commonly accepted that blind source separation is possible because of the statistical independence of separate signal sources. We show that physical restraints of the sources are an equally or an even more important factor when it comes to designing a technical solution for blind signal separation. Based on this new knowledge, we propose a methodology (GeoICA (TM)) for blind source separation by means of an ICA algorithm using components based on geometrical and physical features defined by observation of the signal sources. First, we compare the ICA algorithm in performing blind source separation of a mixture of the female and male speech (signals s1 and s2), and blind source separation of a mixture of computationally generated random signals (signals g1 and g2). Results of our comparison were in compliance with our expectations, because ICA did a much better job (factor > 10) at separating the male and female speech, thus showing importance of physical restraints of the signal sources in blind source separation. Secondly, we propose a methodology (GeoICA (TM)) for real-time blind source separation based on the geometrical and physical features defined by observing the signal sources.