Remedying Sound Source Separation via Azimuth Discrimination and Re-synthesis

Abstract

Commercially recorded music since the 1950s has been mixed down from many input sound sources to a twochannel reproduction of these sources. The effect of this approach is to assign sources to locations in a stereo field using a panposition for each source. The Adress algorithm is a popular way of extracting individual music sound sources from a stereo mixture. A drawback of the Adress algorithm is that when timefrequency components in the stereo mixture are shared between two or more sources, calculating the inter-aural intensity scaling parameter for each source for that time-frequency component is challenging. We show how to obtain a good quality inverse of the pan-mixing process in the time-frequency components which are shared between different sources using a new method called Redress. We demonstrate that we can estimate how much of each source is active in time-frequency components which are shared between sources for two and three-source music mixtures. The consequence of this is that audible artefacts are not as prominent in the source estimates.