An algorithm for spoken keyword spotting via subsequence DTW

We present an algorithm for spoken keyword spotting using subsequence Dynamic Time Warping (DTW) in spoken documents. Instead of using word or phone string as query terms, we use the utterances of user to act as queries. Query matches in the test data are located using subsequence DTW to search between query templates and reference spoken documents. Subsequence DTW is a variant of DTW technique, which is designed to find multiple similar subsequences between two templates. We introduce subsequence DTW into spoken keyword spotting to realize the keyword spotting under low-resource situations in which no in-domain training material is needed. Experiments using this approach are presented using TIMIT corpus.