The accuracy of automatic alignment systems will always be checked using references manually segmented by phonetic or speech communication experts. Consequently, the statistical characterisation of manual segmentation discrepancies becomes quite an important issue. To achieve the goal of finding a better statistical description of these discrepancies, two different experiments were carried out. The first pilot experiment, which was focused on the segmentation of a small subset of an Italian isolated-word speech data-base, was carried out only to better define the second experiment regarding the segmentation of few continuous Italian sentences. This work has been developed in part under the ESPRIT project "Speech Assessment Methodology" (SAM).