Modification in Sequential Dynamic Time Warping for Fast Computation of Query-by-Example Spoken Term Detection Task

Query-by-Example Spoken Term Detection (QbE-STD) under low-resource settings, is the task of retrieval which can be done via the example of an audio. The searching phase involves highly computationally intensive Dynamic TimeWarping (DTW)-based matching techniques. Search space reduction is an important need in order to reduce the space of searching and hence, reduce the computational complexity. In this paper, to perform DTW in a faster mode, the average of consecutive features is considered without overlapping. Much of the information is lost during feature reduction process. For instance, the posterior features on either side of phone boundaries exhibit characteristics. Hence, one such loss might be introduced due to the merging of feature vectors in the vicinity of phoneme boundaries. To overcome this, we perform merging of features after considering the phoneme boundaries (detected using spectral transition measure). The QbE-STD task is performed on MediaEval SWS 2013 dataset. The presented approach reduces the computation time by 46.15% to 49.16 % with very low-performance degradation, i.e., 0.017-0.023 in Maximum Term Weight Value (MTWV) with respect to no feature reduction.