Speech separation based on the images analysis method in CASA

Traditional Computational Auditory Scene Analysis (CASA) method separates speech signal by segmenting and grouping two steps. Combining image analysis techniques acted under correlogram into speech analysis, we propose a novel scheme for these two procedures. A 2-D wavelet transform is employed to segment the speech piths, in order to implement the onset/offset detection. For grouping, we extended a seed region growing method by a new cost function for clustering the image segments of the target speech within segmented speech correlogram. The new approach has been evaluated on mixture speech data and the results demonstrated its efficiency.