Blind separation of mixed audio signals based on improved FastICA

摘要

Independent component analysis has become a predominant method for blind source separation problem. And the FastICA algorithm is widely used due to its rapid convergence property. However, the performance of this algorithm is sensitive to its initial values for the input weight of separation matrix. This paper proposes approaches to improve this algorithm from two aspects. First, performance comparison is made by simulation with three different nonlinear functions used which results in choosing the optimal one. Then the optimal function is further improved by adjusting its parameter. Second, the initial values are calculated with the steepest descent method based on the improved optimal function instead of random choice. The results show that with these techniques we can solve the initial value sensitivity problem, avoid uneven convergence speed and improve the separation effect.