An improvement to the existing blind signal separation (BSS) method has been made in this paper. The proposed method models the inherent signal dependency observed in acoustic object to separate the real-world convolutive sound mixtures. The frequency domain approach requires solving the well known permutation problem, and the problem had been successfully solved by a vector representation of the sources whose multidimensional joint densities have a certain amount of dependency expressed by non-spherical distributions. Especially for speech signals, we observe strong dependencies across neighboring frequency bins and the decrease of those dependencies as the bins become far apart. The non-spherical joint density model proposed in this paper reflects this property of real-world speech signals. Experimental results show the improved performances over the spherical joint density representations.