This study describes how to automatically detect clusters of 2 consonants in a spoken French word recognition system. A set of rules was deduced from the statistic analyses carried out on the 4 corpora recorded by 10 subjects. Five relevant parameters were extracted, namely: 1) voice feature and 2) mode of articulation on the first half of the consonant segment, 3) duration ratio between vowel and consonant segment, 4) duration of consonant segment, and 5) position (prevocalic, intervocalic, postvocalic). This phonetic approach was examined on the GRECO-BDSONS, public test corpus. Our classification was 90% correct using the standard values of the duration parameters provided by our data base.