In this paper the language model for use in multilingual speech technology systems is described that was developed in ESPRIT project 860 'Linguistic Analysis of European Languages'. The model can serve in a variety of applications. In the present contribution attention is focussed on its role in a large vocabulary isolated words speech recognition system. Problems in training and testing a class of models comprising a number of co-operating probabilistic and deterministic knowledge sources are discussed. Results obtained from tests of some of the knowledge sources will be given.