Speech recognisers are normally trained and used by people who can speak words in a predefined vocabulary consistently. People with speech disorders but who make reproducible sounds or utterances can also use a recogniser. However children who experience language and phonological delays, associated with learning difficulties, must be trained to speak clearly and consistently, and the production of templates representing spoken words can be extremely difficult. This paper discusses how templates representing single words can be extracted and edited from a continuous string of utterances made in less than ideal conditions. The edited templates are then used as references for subsequent utterances.