we present a new Time-Frequency approach
for recovering sources’ contribution to two convolutive
mixtures. The separation task is performed on two steps:
Each mixture is clustered into Voiced/Unvoiced frames, and
then the predominant source in each time frequency bin is
identified through a specific weight function which is based
on sources’ excitation characteristics extraction. We investigate
the performance of the proposed approach in the underdetermined
context using objective quality measures.
Results for separating three and four speech sources in a live
recorded mixture show the superiority of the proposed
method in rejecting artifacts over existing convolutive separation
techniques.