This paper is concerned with the definition and description of the phenomenon Off-Talk in human-machine-interaction. This phenomenon is considered to cause problems due to non-relevant information that is conveyed within these utterances. Besides the definition of Off-Talk our work aims to provide an analysis of transcribed audio data that is part of the SmartKom data collection. In the search for features that could indicate the occurrence of Off-Talk we looked at several speech levels e.g. acoustics, lexicon and prosody. Due to the small amount of available data only three features were examined, as there are: loudness, word frequency and filled pauses. The analysis revealed that a correlation might exist between Off-Talk and all features, so that they may serve as indicators for this phenomenon.