Improvements in blind source separation of anechoic underdetermined speech mixtures

The DUET and TIFROM are the most widely used algorithms for separating anechoic mixtures in the underdetermined case (when the number of sources is larger than the number of sensors). Whereas the DUET algorithm requires that the sources are disjoint in the time-frequency domain, the TIFROM considers that there are regions in the time-frequency domain where the mixtures contain only one active source. Since both conditions are not always met in real scenarios, in this work two new algorithms are proposed such as to obtain good estimates of the attenuation and delay coefficients of each mixture without making use of such restrictions. The proposed methods employ binary masking, recursive estimates of the coefficients and sparsifying transforms in order to obtain improved solutions.