On the Identifiability of Latent Models for Dependent Data