Fast algorithm for conditional separation and dereverberation

Last year, we proposed a conditional separation and dereverberation method (the CSD method) for simultaneously achieving blind source separation and dereverberation of sound mixtures. This paper presents a novel fast algorithm for implementing the CSD method. Most of the computation time of the CSD method is spent on calculating what we call modified correlation matrices. The proposed algorithm can calculate these modified correlation matrices much faster than the original algorithm. This improvement is realized by capitalizing on the particular structure of the modified correlation matrices. Experimental results obtained using 672 test samples revealed that the CSD method provided a much better signal-to-interference ratio than a frequency-domain blind source separation method. The real time factor of the proposed algorithm was between 4 and 6, which is less than one tenth that of the original algorithm.