Noise estimation is a fundamental part of speech enhancement. Most traditional methods are heuristic which can not enable the optimal estimation. An unsupervised noise power estimation is presented based on maximum likelihood. A log power statistical model is constructed using hidden Markov model (HMM) in each subband. This model comprises speech and nonspeech Gauss components, and the mean value of nonspeech Gauss component is the estimation of noise power. Moreover, speech may be long term absent, some constraints are introduced to this model for stability. The experiments validate that the proposed method can obtain the maximum likelihood noise estimation and outperforms conventional heuristic methods.