We propose generative modeling algorithms that analyze the temporal features of non-stationary signals and represent their temporal structural dependencies using hierarchical probabilistic graphical models. First, several template sampling methods are introduced to embed the temporal signal features into multiple instantiations of statistical variables. Then the learning schemes that obtain hierarchical probabilistic graphical models from data instantiations are detailed. Based on the sampled temporal instantiations, multiple probabilistic graphical models are discovered and fit to the signal support regions. The evolution structure of these graphical models is depicted using a higher-level structural model. Finally, performance evaluations based on both simulated datasets and audio feature dataset are presented.