Parallel Optimization Of High-throughput Voiceprint Recognition Algorithm For Super-computing Platform

Voiceprint recognition is an important part of biometric identification and is widely used in military intelligence,identification,and retrieval of multimedia information.In today's information explosion,every step of voiceprint recognition faces high throughput and computationally intensive problems.How to use more computing resources to cal-culate massive data while ensuring real-time performance is a key issue that we urgently need to solve now.This paper is based on the super-calculation platform to study the par-allel optimization technology of high throughput rate voiceprint recognition application In this paper,four main modules of decryption,feature extraction,recognition and scoring in voiceprint recognition application are studied.The parallel optimization technology of application on Feiteng many-core processor and MIC multi-core accelerator deployed on Tianhe series supercomputer is studiedAiming at the feature extraction algorithm,a fine-grained time-first optimization algorithm and a SIMD vectorization and OpenMP coarse-grained parallel optimization method are proposed to improve the performance of the feature extraction algorithm Aiming at the voiceprint recognition algorithm based on neural network,the frameless implementation method of ACL function library avoids the performance loss of machine learning framework overhead.A parallel algorithm of voiceprint recognition based on Feiteng processor NUMA architecture is proposed.For the voiceprint information de-cryption process,the parallel optimization of the high throughput decryption algorithm based on the Tianhe super-calculation three-level acceleration framework is designed and implemented.The single super-computing node achieves a speed of 151.95 times com-pared with the single-machine x86 platform.