Word spotting is an appropriate approach to cover real user behaviour , because users often embed target words or phrases in longer utterances than the recognition system expects. In our approach we use a standard Hidden-Markov continuous speech recognition system (CSR) with language model  for word spotting. This word spotting approach does not require any additional HMM-Training for garbage models, because all kinds of garbage models are a concatenation of the standard HMM sub-word-units of the system. As we want to use our word spotting system as an open input, we principally do not constrain the algorithm to only one keyword or phrase per utterance. It is also shown that there is a soft transition between word spotting, phrase spotting and a standard CSR.
Bibliographic reference. Klemm, H. / Class, Fritz / Kilian, Ute (1995): "Word- and phrase spotting with syllable-based garbage modelling", In EUROSPEECH-1995, 2157-2160.