This thesis deals with the problem of source separation with main focus on music signals. Source separation algorithms aims at separating individual sources from a mixture. These type of algorithms are generally used to solve problem known as the "Cocktail party effect" where algorithms are trying to extract the sources "blindly". The blindness of these algorithms comes from the unknown mixing parameters of the sources which were used for creating mixtures. The current techniques use statistical and mathematical methods for source separation that do not take any advantage of the signal characteristics of source. This thesis presents a novel approach to source separation which uses signal characteristics to separate source.