Improving model-based convolutive blind source separation techniques via bootstrap

Blind source separation for underdetermined reverberant mixtures is often achieved by assuming a statistical model for cues of interest where the unknown parameters of the statistical model depend on hidden variables. Here, the expectation-maximization (EM) algorithm is employed to compute maximum-likelihood estimates of the unknown model parameters. A by-product of the EM algorithm is a time-frequency (T-F) mask which allows the estimation of the target source from the given mixture. In this paper, we propose the idea of bootstrap averaging to improve separation quality from mixtures recorded under reverberant conditions. Our experiments on real speech mixture signals show an increase in the signal-to-distortion ratio (SDR) over a state-of-the-art baseline algorithm, to our knowledge, currently, the best performing technique in this class of methods.