Nonstationary source separation for underdetermined speech mixtures

We propose a multichannel source separation method for underdetermined mixtures of nonstationary signals, such as speech. Like other underdetermined algorithms, our method relies on the time-frequency sparsity of speech. However, our interference model allows more than one source to be active at the same time and frequency, providing better separation performance for mixtures of many sources. The system consists of several beamformers designed for different combinations of interference sources. A decision rule selects the beamformer that best suppresses the active interferers at each time-frequency point. Experiments on both simulated and real mixtures show improved interference suppression compared to conventional beamformers.