We present a reallistic and robust implementation of Blind Source Separation and Blind Deconvolution. The algorithm is developed from the idea of natraul gradient learning, wavlet filtering and denoising, and the characteristic of different sound source. Several hardware pieciecs are integrated, including a mobile robot, NT workstation and DSP chip to achieve the real time separation of real world signal. Besides, a method of judging the separation performance without knowing the mixing matrix ( mixing filter ) is proposed and verified.