An Improved Speech Blind Separation Algorithm Based on Non-Linear Function

Abstract

This paper introduces a new algorithm based on non-linear function to adaptively control step-size which is used for updating separation matrix to extract a target speech source accurately in blind source separation (BSS). The use of fixed step-size parameter of the conventional BSS algorithm usually results in a trade-off between convergence speed and steady-state misadjustment. The presented algorithm will eliminate much of this trade-off. It intelligently regulates the step-size according to the time-varying dynamics of other parameters at each iteration. The desirable ability of the new algorithm to improve convergence speed and steady-state misadjustment is demonstrated by MATLAB simulation results.