Skip to Main Content
Gradient method because of its good convergence speed and separation performance, it is an important way for blind source separation. But in order to offset the expansion factor of uncertainty, the basic natural gradient algorithm forces sources to have constant magnitude. When the source are nonstationary signals, such constraints forced rapid changes in rate of separation matrix, which will affect the numerical stability. To avoid this drawback, a variable step-size method proposed based on the natural gradient algorithm with non-holonomic constraints (N-VS-NG). Experimental results verify the superior convergence performance over conventional natural gradient (NG) in nonstationary environments.