Sentence hypothesisation in a speech recognition and understanding dialogue system for the Slovenian language is presented. A statistical approach using the 3g-gram model based on Jelinek's trigram model was applied to the task of linking word hypotheses together into grammatically well-formed sentences. The equivalence classes are given by 72 different parts of speech, extended by various morpho-syntactic features. A simple smoothing method is presented and compared to a standard one.