A parallel computing approach to CNN-based QbE-STD using kernel-based matching.

In query-by-example spoken term detection (QbE-STD), reference utterances are matched with an audio query. A matching matrix-based approach to QbE-STD needs to compute a matching matrix between a query and reference utterance using an appropriate similarity metric. Recent approaches use kernel-based matching to compute this matching matrix. The matching matrices are converted to grayscale images and given to a CNN-based classifier. In this work, we propose to speed up QbE-STD by computing the matching matrix in parallel using a coarse-grained data parallelism approach. We explore two approaches to coarse-grained data parallelism: In the first approach, we compute parts of the matching matrix in parallel and then combine them to form a matching matrix, while in the second one, we propose to compute matrices in parallel. We also propose to convert the matching matrices into two-colored images using the threshold and use these images for QbE-STD. The efficacy of the proposed parallel computation approach is explored using the TIMIT dataset. [ABSTRACT FROM AUTHOR]