A lack of suitable databases limits development of general large vocabulary continuous speech recognition systems for Norwegian speech, forcing researchers to look at easier tasks. One such task is dictation of radiologic diagnoses. Even though the total vocabulary is very large, each doctor uses a much more limited vocabulary for each specific examination type. This paper de- scribes the development of a speaker dependent system with a 977- word vocabulary for one such examination type. The Hidden Markov Model Toolkit (HTK) from Entropic was used to train context-in- dependent phoneme models and a bigram language model. By reducing the phoneme set, using a bigram probability floor and letting the phoneme-specific number of Gaussian mixture elements depend on the amount of available training data, promising recognition re- sults were obtained with limited amounts of training data. The development process was considerably simplified by using general speaker independent acoustic models for initiation of our models, and by letting the text-to-phoneme part of a Norwegian system for speech synthesis provide the phonemic spelling of the vocabulary.