In this article, we propose an on-line algorithm for Blind Source Separation of speech signals, which is recorded in a real environment. This on-line algorithm makes it possible to trace the changing environment. The idea is to apply some on-line algorithm in the time-frequency domain. We show some results of experiments.