Generative models can be flexibly employed in a variety of tasks such as classification, detection and segmentation thanks to their explicit modelling of likelihood functions. However, likelihood functions are hard to model accurately in many real cases. In this paper, we present an enhanced hidden Markov model capable of dealing with the noisy, high-dimensional and sparse measurements typical of action feature sets. The modified model, named hidden Markov model with multiple, independent observations (HMM-MIO), joins: a) robustness to observation outliers, b) dimensionality reduction, and c) processing of sparse observations. In the paper, a set of experimental results over the Weizmann and KTH datasets shows that this model can be tuned to achieve classification accuracy comparable to that of discriminative classifiers. While discriminative approaches remain the natural choice for classification tasks, our results prove that likelihoods, too, can be modelled to a high level of accuracy. In the near future, we plan extension of HMM-MIO along the lines of infinite Markov models and its integration into a switching model for continuous human action recognition.