基于粒子对协同优化的说话人辨认

Speaker Identification Based on Particle-Pair Cooperative Optimizer

本文提出了一种新的说话人码本的优化设计方法—粒子对协同优化算法,应用于矢量量化的说话人辨认.此算法利用两个初始粒子对分别在每次迭代中执行粒子群优化算法的速度、位置更新和标准LBG算法实现并行搜索最优码本,粒子对由两个粒子构成,每隔一定的迭代次数通过交换粒子实现粒子对间的信息交流,最后分别选出两个较优粒子组成精英粒子对进一步搜索.此算法避免传统LBG算法陷入局部最优的缺点.实验结果表明,本算法始终稳定地取得显著优于LBG、FCM、FRLVQ-FVQ、FEP和PSO算法的说话人辨认性能,较好地解决了初始码本影响优化结果的问题,且在计算时间和收敛速度方面有优势.

A novel particle-pair cooperative optimizer（ PPCO） is proposed for speaker identification based on vector quantization. In this algorithm, two initial particle-pairs simultaneously explore for the best speaker codebook, and each particle-pair which consists of two particles performs basic operations of particle swarm optimization（ velocity updating and position updating）and conventional LBG algorithm in sequence at each iteration. Information is exchanged when particle-pairs are reorganized periodically. And then two elitist particles selected from two initial particle-pairs respectively continue to move toward the global optimum. Experimental results have demonstrated that the performance of this new algorithm is much better than that of LBG, FCM, FRLVQ- FVQ,FEP and PSO consistently with lower speaker identification error rates, shorter computational time and higher convergence rate. The dependence of the final codebook on the selection of the initial codebook is also reduced effectively.