A Novel Bi-Group Particle Swarm Optimizer for Speaker Identification

Abstract

The traditional training meshods of speaker codebook for speaker identification based on vector quantization are sensitive to the initial codebook parameters, and they often lead to a sub-optimal codebook in practice. To resolve this problem, this paper proposes a novel bi-group particle swarm optimizer (BPSO). It applies two sub-group particles with different particle update parameters simultaneously to explore the best speaker codebook, and the particles perform basic operations of particle swarm optimization (PSO) and conventional LBG algorithm in sequence, which can explore the solution space separately and search the local part in detail together. Information is exchanged when sub-groups are periodically shuffled and reorganized. Experimental results have demonstrated that the performance of BPSO is much better than that of LBG, fuzzy C-means (FCM), fast evolutionary programming (FEP), PSO, the impoved PSO algorithm consistently with higher correct identification rates and convergence rate. The dependence of the final codebook on the selection of the initial codebook is also reduced effectively.