The success of speaker-dependent speech recognition systems will largely depend upon the consistency between utterances made during reference template training and those made during system application. Inconsistencies can result from differences in the environmental conditions, the behaviour of the user or because of the discrepancy between task demands made during the application compared to template training. A careful scheme of user induction can help, but the aim for designers must to prepare system procedures that help utterance consistency. Suggestions are made for doing this by improving methods of harvesting utterances on which reference templates will be based.