混合粒子对优化算法在说话人识别中的应用

A Novel Shuffled Particle-pair Optimizer for Speaker Recognition

在粒子群优化(Particle Swarm Optimization,PSO)和混合蛙跳算法(Shuffled Frog-Leaping Algorithm,SFLA)的基础上,该文提出了一种新的混合粒子对优化(Shuffled Particle-Pair Optimizer,SPPO)算法,应用于矢量量化的说话人识别。该算法将全局信息交换和局部深度搜索相结合寻求最佳的说话人码本。群体按适应值分为3个粒子对,每个粒子对由两个粒子构成,按先后顺序执行PSO算法中的速度位置更新和LBG算法以实现局部细致搜索,间隔一定的迭代次数通过SFLA混合策略实现粒子对间的信息交换,从而使群体向全局最优解靠近。实验结果表明,本算法始终稳定地取得显著优于LBG,FCM,FRLVQ-FVQ和PSO算法的说话人识别性能,较好地解决了初始码本影响的识别性能的问题,且在计算时间和收敛速度方面有相当的优势。

A novel Shuffled Particle-Pair Optimizer （SPPO） is proposed for speaker recognition based on vector quantization, which combines the advantage both in Particle Swarm Optimization （PSO） and Shuffled Frog-Leaping Algorithm （SFLA）. The SPPO contains elements of local exploration and global information exchange to get global optimized speaker codebook. In this algorithm, the population is partitioned into 3 particle-pairs according to the performance, and each particle-pair consists of two particles. The particle-pairs perform simultaneously local exploration using basic operations of PSO （velocity updating and position updating） and LBG algorithm in sequence. A shufflingstrategy, in which the particles are periodically shuffled and reorganized into new particle-pairs, allows for the exchange of information between particle-pairs to move toward the global optimum. Experimental results demonstrat that the performance of this new method is much better than that of LBG, FCM, FRLVQ-FVQ, and PSO consistently with lower speaker recognition 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.