This paper proposes an original methodology devoted to the automatic detection of abnormal zones in speech utterances in the specific context of impairments. This methodology relies on automatic speech processing, involving an automatic text-constrained phoneme alignment, the computation of phoneme based-normalized acoustic scores and a reference scale, permitting to label in fine a phoneme as normal or abnormal. The evaluation of the methodology reliability when applied to a dysarthric speech corpus has shown very encouraging results, highlighting the efficiency of the methodology in detecting true abnormal zones. In addition, this evaluation underlines a lack of precision (in terms of " information retrieval ") resulting in mis-labelled normal zones in the automatic annotation (false positive), leading to further investigation.