Be more specific describing your problem.
Voice recognition and voice separation are very different problems.
In the last case, you could use the ICA (Independent Component Analysis) approach. There is a free matlab-based toolbox called FastICA. You can find it in google. With this technique you will need as many signals as sources (at least).
Another techniques are BSS (blind source separation) and BSD (blind source deconvolution), which is necesary when the voices are recorded in a room with high reverberation.