The predominant learning strategy for H(S)MMs is local search heuristics, of which the Baum-Welch/ expectation maximization (EM) algorithm is mostly used. It is an iterative learning procedure starting with a predefined topology and randomly-chosen initial parameters. However, state-of-the-art approaches based on arbitrarily defined state numbers and parameters can cause the risk of falling into a local optima and a low convergence speed with enormous number of iterations in learning which is computationally expensive. For models with persistent states, i.e. states with high self-transition probabilities, we propose a segmentation-based identification approach used as a pre-identification step to approximately estimate parameters based on segmentation and clustering techniques. The identified parameters serve as input of the Baum-Welch algorithm. Moreover, the proposed approach identifies automatically the state numbers. Experimental results conducted on both synthetic and real data show that the segmentation-based identification approach can identify H(S)MMs more accurately and faster than the current Baum-Welch algorithm.