Blind separation of sources consists of recovering a set of signals in which only instantaneous linear mixing are observed.This paper presents a novel blind separation method of nonstationary sources when noises are

α

-stable processes

the observed data are preprocessed using an empirical threshold value.To make full use of nonstationarity and temporal correlation of sources

we exploit sliding windows to form multiple time-delay correlation matrices of the weighted observed data

then use approximately joint diagonalization to estimate the mixing matrix and source signals.The method is limited to the case of characteristic exponent

α

tending to unity

and computer simulation is provided to illuminate the high performance of the proposed method.