Single-channel speech separation using sequential discriminative dictionary learning

A novel sequential discriminative dictionary learning (SDDL) algorithm is presented to suppress the confusion between the separated signals which we denote source confusion. The existing discriminative dictionary learning (DDL) algorithms assume that signals from different speakers have their unique components which makes that the signals can be explained by the corresponding sub-dictionaries. But the signals from different speakers have similar components when divided into speech segments. We take the unique and similar components of different speakers׳ signals into account, and design a new structured dictionary which contains discriminative and buffer sub-dictionaries. The unique components of different speakers׳ signals which have better correspondences to the speakers׳ labels are firstly separated, and the similar components of different speakers׳ signals are separated in the next layer. An objective function is derived, which guarantees that the unique components of the training sets can be explained by their corresponding discriminative sub-dictionaries and the similar components of the training sets can be explained by the buffer sub-dictionary rather than the cross sub-dictionary. The components distributed in the buffer sub-dictionary are used as training sets in the next layer. Experiments results verify that the proposed algorithm can effectively reduce the source confusion compared to the existing algorithms.