We discuss on a word spotting strategy that incorporates a language model as heuristics. It judges the word existence based on not only the matching score of the word itself but also the plausibility of the rest part as a sentence. Several language models are examined with respect to the accuracy and robustness, and the followings are conclusions: (1) Syllable-level knowledge is robust but insufficient. (2) Word-level knowledge is very effective and robust against spontaneous speech. (3) Word-pair constraint is most powerful but not robust. We further propose to incorporate phrase-level syntax into the spotting unit. The phrase-level syntax is rarely violated even in spontaneous utterances and significantly reduces the task perplexity. It turned out effective especially in getting a higher detection rate with a small number of false alarms.