Under-determined reverberant blind source separation using excitation sources characteristics

This paper presents a new blind source separation technique, separating more than two sources from a stereo recording mixture. It is based on time delay of arrival (TDOA), and instants of significant excitation (GCI´s) of each speaker. The source separation task is achieved by determining a weight function specific to each speaker to process the linear prediction (LP) residual of the observed mixed speech. Each separated speaker´s speech is synthesized by exciting an all pole filter by its corresponding processed LP residual. The Algorithm performance is evaluated on reverberated mixtures using objective quality measures. Promising results are obtained when comparing the proposed approach with other separation techniques in the field.