A new speech enhancement technique is proposed assuming that a speech signal is represented in terms of a linear probabilistic process and that a noise signal is represented in terms of a stationary random process. Since the target signal, i.e., speech, cannot be represented by a stationary random process, a Wiener filter does not yield an optimum solution to this problem regarding the minimum mean variance. Instead, a Kalman filter may provide a suitable solution in this case. In the Kalman filter, a signal is represented as a sequence of varying state vectors, and the transition is dominated by transition matrices. Our proposal is to construct the state vectors as well as the transition matrices based on time-frequency pattern of signals calculated by a wavelet transformation (WT). Computer simulations verify that the proposed technique has a high potential to suppress noise signals.