In this paper, we consider the case of single microphone Blind speech separation. We exploit the joint model of speech signal (the voiced part) that consists on modeling the correlation of speech with a short term Autoregressive process and its quasi-periodicity with a long term one. a linear state space model with unknown parameters is derived. The separation is achieved by estimating the state as well as the unknown parameters. This task is assured by the use of Kalman filtering modified with the variational bayes techniques witch takes into consideration the estimation error of parameters used in Kalman filter.