Real-time Blind Separation and Deconvolution of Real-world signals

摘要

We present a realistic and robust implementation of Blind Source Separation and Blind Deconvolution. The algorithm is developed from the idea of natural gradient learning, wavelet filtering and denoising, and the characteristic of different sound source. Several hardware pieces 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.