A flexible spatial blind source extraction framework for robust speech recognition in noisy environments