We propose an improvement to the centisecond TLHMM model (Meziane, 1999) applied to the sound duration. Indeed, the distribution of the sound duration depends on the speaking rate. An adaptation in a post-processing step is needed. This adaptation is studied by proposing a model of the speaking rate based on average syllabic duration. The experiments elaborated on a set of BDSONS show the interest of this approach. This work is a continuation of those by Meziane (1999) and Suaudeau (1994).