For many applications it is not necessary or even desirable to recognise every word in an utterance. A more satisfactory approach may be to try to identify the occurrence of a number of key words within the utterance string. In this paper we describe a system which identifies key words occurring in unconstrained conversation.