In this paper, we investigate a weighted finite-state transducer approach to language modelling for speech recognition applications. We explore a unified framework to conversational speech recognition which combines the benefits of grammars, n-gram and class-based language models, with the flexibility of using dynamic data, and the potential for integrating semantics. Based on a virtual personal assistant application, we show first applications and recognition results of out-of-grammar handling and the integration of class-based, weighted, dynamic data into this framework.