The present study, which reports the first stage of analysis of five corpora containing a total of 1488 errors, was aimed firstly at classifying French speech errors and determining whether our results are consistent with those obtained by most authors; secondly, by restricting our analysis to the phonological level, we have tried to discover the underlying mechanisms of phonologically-based lapsus linguae. Most of the results obtained are in line with those of previous studies. The results on phonological errors thus corroborate Shattuck-Hufnagels (1980) availability and similarity theory. However the results on vowel errors showed in particular that speech errors are not confusions and cannot be compared to phonological neutralizations, given that the similarity principle must be based on robust features.

Accurate analysis of the feature substitution hierarchy showed that a permutation graph whose axes are the dimensions of the vocal tract is the best model for predicting speech errors.