Research On Heterogeneous Parallel Computing For Speaker Recognition System Based On Big Data

Speaker recognition is an important branch of speech recognition and one of thehotspot. Each step included in speaker recognition system requires high intensivemathematical calculations on a large number of files. It will take a long time forspeaker recognition based on massive data perform on one single computer. GPU hasbecome the leader of promoting the floating-point performance, using NVIDIACUDA-C can help us program on GPU without an advanced understanding ofgraphics algorithm. This paper studies an parallel programming mode based onspeaker recognition system with big data.Firstly, the paper introduces the speaker recognition system and parallelcomputing model, analyzes the disadvantages of speaker recognition system based onmassive data. It leads to the discussion of concurrent programming and analyzes theresearch condition on the parallel programming both domestic and foreign.Secondly, this paper proposes a set of design plan of using both multi-core CPUand massive-core GPU on speaker recognition system. We focus on the study ofCUDA parallel computing model of GPU and CPU multi core calculation modelbased on Python programming language. From the top to down, this paper designs amulti processes queue model, it is an scheduling model used in CPU multiple cores, itdrives worker processes based on multi threads in GPU massive cores. Above all, thispaper presents a complete heterogeneous calculation method based on speakerrecognition system using GMM-UBM model.Finally, the whole system is tested to verify the stability and time-consumedperformance. We tested the speaker recognition system’s recall rate both on serial andparallel structure first, then we got objective testing data through the verification ofconcurrent performance of each module. At the end of this paper, we conduct thistopic research and put forward a direction of future development of this subject.