We introduce in this paper a hierarchical hybrid statistical language model, represented as a collection of local models plus a general model that binds together the local ones. The model provides a unified framework for modelling language both above and below the word level, and we exemplify with models of both kinds for a large vocabulary task domain. To our knowledge this is the first paper to report an extensive evaluation of the improvements achieved from the use of local models within a hierarchical framework in comparison with a conventional word-based trigram model.