Highly scalable speech processing on data stream management system

Abstract

Today we require sophisticated speech processing technologies that process massive speech data simultaneously. In this paper we describe the implementation and evaluation of a Julius-backended parallel and scalable speech recognition system on the data stream management system "System S" developed by IBM Research. Our experimental result on our parallel and distributed environment with 4 nodes and 16 cores shows that the throughput can be significantly increased by a factor of 13.8 when compared with that on a single core. We also demonstrate that the beam management module in our system can keep throughput and recognition accuracy with varying input data rate. © 2012 Springer-Verlag.