In a speech production model proposed by Levelt a distinction is made between two routes of phonetic implementation in speech. A syllabary route is used to retrieve the stored motor programs for the most frequent syllables of a language, and segment-by-segment assembly is used for the implementation of low-frequency syllables. One of the predictions of the model is that there should be a difference in coarticulation between motor programs retrieved from the syllabary and programs that are computed online. In this paper we present two laboratory experiments and a corpus study on German which were designed to verify this prediction. Our results support the hypothesis that articulatory programs for high-frequency syllables are implemented differently than those for rare syllables.