This work deals with the phenomenon of vowel reduction in spontaneous speech. Automatic and manual analyses have been conducted on a large conversational speech corpus (CID) to study the extra-short vowels (less than 30 ms), generally excluded from the automatic analyses. A strong reduction in the vocalic system and very short durations are observed for a great proportion of vowels in the corpus. A manual analysis highlights the specific realisations of these extra-short vowels: they are more often in function words than in content words; almost all of them belong to monosyllabic words; their formant values show strong dispersion in F1/F2 plan; a large context is needed for their identification.