Use of GPU and Feature Reduction for Fast Query-by-Example Spoken Term Detection

For query-by-example spoken term detection (QbE-STD) on low resource languages, variants of dynamic time warping techniques (DTW) are used. However, DTW-based techniques are slow and thus a limitation to search in large spoken audio databases. In order to enable fast search in large databases, we exploit the use of intensive parallel computations of the graphical processing units (GPUs). In this paper, we use a GPU to improve the search speed of a DTW variant by parallelizing the distance computation between the Gaussian posteriorgrams of spoken query and spoken audio. We also use a faster method of searching by averaging the successive Gaussian posteriorgrams to reduce the length of the spoken audio and the spoken query. The results indicate an improvement of about 100x with a marginal drop in search performance.