A bottom-up phonetic HMM recognizer will be combined with a long-span template based recognizer, with the aim to outperform a traditional HMM recognizer. A novelty is the way in which the phoneme graphs obtained for training data will be used to learn alternative transcriptions of words and Multi-word expressions and to incorporate them in the lexicon. On top of that, the template-based recognizer will only activate templates that appear in one of the learned variants.