Keyword spotting based on mixed grammar model

We present a novel keyword spotting method based on the mixed grammar model. By merging the filler model and the finite state grammar, two conventional technologies of keyword spotting, the mixed grammar model incorporates both a priori knowledge and the capability of covering all possible sentential forms in real speech, thus makes up for the weaknesses of both parental technologies. Experimental results show that the mixed grammar model excels the filler model in overall performance and the finite state grammar in robustness. The expansibility of the mixed grammar model is shown in its capacity of easy incorporation of further improvement of both the filler model and finite state grammar.