Abstract: Considering the number of uncertainty in dynamic speech signal's blind sources separation on particular occasions, expanding fourth-order cumulant matrix to estimate adaptively the number of speech signal sources is proposed. The method overcomes the shortage of traditional blind sources separation which presupposes the speech signal source number has been confirmed and adaptively chooses the corresponding blind sources separation algorithm to realize speech signal's separation only according to the relationship between number of observing speech signal and source speech signal. Simulation experimental results show that the algorithm can effectively realize overdetermind, posed-determind and underdetermind speech signal's separation with good separation performance and robust.