Spoken term detection search into acoustic documents for specific acoustic patterns. In our query by example approach We focus in a totally unsupervised search of an uttered acoustic pattern inside a bigger acoustic document. We present a second level approach for building an unsupervised acoustic model. This acoustic model is used to represent and effectively compare acoustic frames. Posteriorly, we use segmental dynamic time warping for subsequence matching. Resulting alignments are normalized and filtered to produce the final putative hits.