This paper proposes a method to estimate the spectrographic speech mask based on a two-dimensional (2-D) correlation model. The proposed method is motivated by a fact that the time and frequency correlations of speech presence are interwoven with each other in the time-frequency (TF) domain. Conventional Markov chain is incapable of simultaneously modeling the time and frequency correlations in an adaptive way. The 2-D correlation model is presented to describe the correlation of speech presence in the TF domain, where the speech presence and absence are taken as two states of the model. The time correlation is modeled by the time state-transition probability and the forward factor, while the frequency state-transition probability and the corresponding neighbor factor are defined to describe the frequency correlation. The time and frequency correlations are incorporated into the model by maximizing the Q-function. A sequential scheme is presented to online estimate the parameter set. Given the observed spectrum and the parameter set, the state matrix that maximizes the posteriori probability is regarded as the optimal estimate of the speech mask. The proposed method was compared with some well-established methods. The experimental results confirmed its superiority.