It was established that the performance of our annotation system [8] is affected by the length of the utterances: the error rate, the CPU-load and the memory requirements tend to increase as the utterances get longer. In this contribution the speech signal is first segmented into speech, pauses and noise (breaths, clicks, : : :) and subsequently split in signal phrases prior to the annotation. Experiments on 3 different databases (3 languages) demonstrate that this stategy yields a significant improvement of the annotation accuracy.