The investigation presented in this paper is a first attempt to specify the reliability of manual segmentation and labelling. It is demonstrated that transcription performance does not only rely on external conditions, i. e. segmentation criteria, transcribers' labelling experience or characteristics of labelling devices but rather depends on internal acoustic-phonetic features of the analysed utterance and the level of abstraction of description. Three levels of transcription are introduced. The reliability of segmented data at each level is discussed on the basis of consistency measures within the judgements of different transcribers.

Keywords: Speech segmentation, labelling, levels of transcription, interindividual consistency