This paper presents a new nonstationary HMM that resolves some of the complexities of semi-Markov models and the lack of temporal coherency in the previous nonstationary HMM. Unlike the semi-Markov model, it is fully first-order Markovian. But the temporal evolution of the transition distribution renders it effectively a high-order model. In the proposed model, once a Markov chain enters a state, it tends to stay in the state for a while but gradually the tendency of exiting the state grows. This tendency is described by a gamma CDF(cumulative distribution function) and the resulting model behavior is highly intuitive. Using the evaluation and training algorithms as provided here, a set of handwriting digit models were trained and analyzed in three ways including classification with comparable performance to that of semi-Markov models, and random sampling with the least distortion and highest quality. The proposed model is believed to have great advantages in accurate and detailed discrimination of ambiguous patterns, and pattern generation and synthesis applications.