关键词：自动语音识别；ASR；应用程序
摘 要：Data layout, data placement, and synchronization processes are notusually part of a speech application expert’s daily concerns. Yetfailure to carefully take these concerns into account in a highly parallelimplementation on the graphics processing units (GPUs) couldmean an order of magnitude of loss in application performance. Inthis paper we present an application framework for parallel programming of automatic speech recognition (ASR) applications thatallows a speech application expert to effectively implement speechapplications on the GPU. It is an approach for crystallizing and transferring the often tacit knowledge of effective parallel programmingtechniques while allowing for flexible adaptation to variousapplication usage scenarios.