High fidelity blind source separation of speech signals

This paper addresses blind source separation (BSS) problem of multiple speech signals in low signal-to-interference-noise ratio (SINR) environment. We consider an over-determined case so that we can form multiple sub-arrays (of which there are as many sensors as speech signals), and propose a novel hybrid scheme to obtain high fidelity speech signals after separation. Firstly, the proposed method applies the commonly-used BSS technique at each sub-array to separate the speech signals. Next, the outputs of the same speech signal from different sub-arrays are grouped to form a new sub-array. We can then exploit the spatial diversity of the new sub-array to achieve high fidelity source separation. This configuration is the key innovation of this paper. Another contribution of the paper is on the justification of using the hybrid configuration to further increase the output SINR. From numerical analysis, it is demonstrated that 12 dB SINR improvement can be achieved using 5-element sensor array in the presence of two other interfering speech signals over a range of input SNR values. A significant improvement can also be seen from the output signal-to-artifact ratio (SAR) of the recovered signals.