The transition from high to low mortality levels have been accompanied by two apparent paradoxes: The convergence and the deceleration of the mortality rates at old ages. On one hand, while death rates at all ages have been dramatically reduced over time, the rate of ageing has increased considerably (convergence). On the other hand, adult and early-old mortality increases exponentially up to about age 80 and thereafter slows down (deceleration). Both of the above evidences fit with the theory of the selective survival. Several parametric and non parametric models incorporating period and cohort components have been devised and applied to long series of mortality data over age and time, without however bringing to conclusive results. In this paper, we (try to) shed light on the mortality dynamics over the long term by applying a hidden Markov approach to analyse mortality surfaces over age and time. The idea is to cluster death rates over age and time according to a (given) number of latent classes, by accounting for the temporal correlation between death rates. Each latent class is associated with a Gompertz model. The transitions between latent classes are driven by a Markov chain that segments the mortality surface.