Speech Recognition on Modern Graphic Processing Units
Department of Electrical and Computer Engineering, Northeastern University, Boston, MA
Sixth Annual Boston Area Architecture Workshop, 2015
@article{yu2015speech,
title={Speech Recognition on Modern Graphic Processing Units},
author={Yu, Leiming and Magrath, John and Pandey, Ajey and Sears, Matthew and Kaeli, David},
year={2015}
}
Speech Recognition run on Graphic Processing Units (GPUs) has shown some promising performance improvements ranging 2-10x speedups when compare to execution on CPUs. GPU has continued to introduce new programming features, such as Dynamic Parallelism and Hyper-Q, that could further benefit Speech Recognition processing. In this paper we describe a framework developed at Northeastern describing our Speech Recognition system NUSpeech. In this system, processing is dominated by processing in a Hidden Markov Model (HMM). Leveraging the computation power of GPUs, we have been able acceleration the processing of the HMM by 9.2x. Not limited to discrete GPUs, our implementations are portable to GPU-integrated embedded systems. Our goal is further accelerating NUSpeech to perform real-time speech processing while sustaining a high energy efficiency.
February 9, 2015 by hgpu
