Large vocabulary continuous speech recognition is a very hard task, which is the reason why state of the art systems use multiple sources of linguistic knowledge to better solve the problem. Those knowledge sources refer to various modelling levels, such as the acoustic, phonetic, lexical and syntactic levels. Due to the diversity of knowledge sources which must be integrated, algorithms for recognizing large vocabulary continuous speech are traditionally very complex. The integration of new knowledge sources can be problematic, since each new knowledge source adds to the complexity of the recognizing algorithm. The INESC-ID speech recognition system addresses this complexity through the use of Weighted Finite-state Transducers (WFST). WFSTs are finite-state machines that allow modelling weights relations and provide a unifying formalism for speech recognition. In this talk, the INESC-ID WFST approach to automatic speech recognition will be presented. A state of the art system for transcribing broadcast news in European Portuguese will be demonstrated. It will also be shown how WFST techniques are currently being applied to speech-to-speech machine translation.