An algorithm for single channel signal separation is presented. The algorithm projects the observed signal to given subspaces, and recovers the original sources by probabilistic weighting and recombining the subspace signals. The results of separating mixtures of two different natural sounds are reported.