Blind separation of speech with a switched sparsity and temporal criteria

A blind signal separation algorithm (SCAtemp) that exploits both the sparse time-frequency representation and temporal structure of speech is proposed. SCAtemp compares each speech signal's adherence to the sparsity and temporal criteria, before switching to the most appropriate criteria to estimate each signal. This algorithm is shown to improve the real time separation performance of conventional BSS algorithms exclusively exploiting either the temporal structure, sparsity or statistical independence of signals. The improvement of SCAtemp over conventional BSS algorithms can be attributed to the use of additional a priori knowledge of speech in the temporal short term.