Online blind audio source separation using recursive expectation-maximization

The challenging problem of online multi-microphone blind audio source separation (BASS) in noisy environment is addressed in this paper. We present a sequential, non-iterative, algorithm based on the recursive EM (REM) framework. In the proposed algorithm, the compete-data, which constitutes the separated sources and residual noise, is estimated in the E-step by applying a multichannel Wiener filter (MCWF); and the corresponding parameters, comprised of acoustic transfer functions (ATFs) relating the sources and the microphones and power spectral densities (PSDs) of the desired sources, are sequentially estimated in the M-step. The separated speech signals are further enhanced using matched-filter beamformers. The performance of the algorithm is demonstrated in terms of the separation capabilities, the resulting speech intelligibility and the ability to track the direction of arrival (DOA) of the moving sources.