This paper investigates the quality of the automatic phonetic segmentation of non-native speech with respect to the amount of pronunciation variants that are used in the pronunciation lexicon, in view of exploiting such segmentation for providing automatic prosodic feedback for computer assisted foreign language learning. The selection of the pronunciation variants in the lexicon was controlled according to the amount of times the variants were observed in non-native speech from other speakers. Keeping all the variants that were observed at least one time in other non-native speaker data, leads to the worst results, confirming the fact that introducing unusual pronunciation variants in the lexicon is also harmful for non-native speech segmentation. Overall best results are obtained when keeping all the variants observed at least two times. This seems a good trade-off to have a large number of variants, and at the same time avoiding unusual variants.