This paper presents an approach to automatically extract and evaluate the ``stability'' of pronunciation variants (i.e., adequacy of the model to accommodate this variability), based on multiple pronunciations of each lexicon words and the knowledge of a reference baseform pronunciation. Most approaches toward modelling pronunciation variability in speech recognition are based on the inference (through an ergodic HMM model) of a pronunciation graph (including all pronunciation variants), usually followed by a smoothing (e.g., Bayesian) of the resulting graph.