We describe a new approach for generating Norwegian intonation in text to speech synthesis. The method is based on a phonological representation of utterances. The overall f0 contour of an utterance is synthesised by concatenation of stored f0 contours corresponding to accent units. Candidate accent units are found by searching a lexicon derived from natural speech and selecting the unit that is the best match with respect to the properties of the target accent units of the utterance to be synthesised. A formal subjective test confirms that the new approach leads to more natural speech than a former rule based method, but the quality is still inferior to intonation copied from natural speech.