For the the development of industrial speech recognition systems a set of phonemically balanced sentences read by six German speakers has been manually segmented and labelled by a group of well trained transcribers. In addition three subjects re-labelled the set of utterances from one speaker 10-12 months after the first session. This paper presents an analysis of labelling results in terms of intra- and interindividual consistency using Prolog tools for detailed examination of our data. We show that inconsistency in judgements is not equally distributed over the whole range of phonetic categories but rather is concentrated on a smaller subset of critical cases. Moreover, intra- and interindividual agreement between labelled series of utterances turned out to be speaker specific and highly dependent on phonetic context. The consequences of our results for an optimization of manual labelling procedures are discussed.