Hidden Markov Models (HMMs) are now widely used for isolated word and continuous speech recognition and, given their success, have been applied recently to the keyword spotting (KWS) problem. In this paper, a new approach is presented, which does not attempt to explicitly model extraneous speech. This approach is tested on a speaker independent, over the telephone line, 10 word lexicon database and compared with several other KWS algorithms using explicit garbage modelling. As the resulting methods should be used in a real time system, we focus on these approaches that do not require extensive computation.