Explaining the effects of labial coarticulation is a difficult problem that gave rise to many studies and models. Most of the time, small corpora were exploited to design these models. In this paper we describe the realization and exploitation of a corpus with ten speakers. This corpus enabls the most invariant labial features (protrusion, stretching and lip opening) to be established. Then we propose a formal prediction algorithm that relies on a standard phonetic description of french phonemes. We conducte a first evaluation of this algorithm that shows its relevancy.