基于遗传算法的矢量量化的话者模型

SPEAKER MODEL BASED ON GAVQ ALGORITHM

本文提出了基于遗传算法(GA)的矢量量化话者模型,简称GAVQ话者模型。给出了建立该模型的算法即GAVQ算法。矢量量化技术可以有效地提取和表征话者的个性信息,因而VQ码本可以有效地用作话者模型,传统的VQ码本设计方法是LBG算法,但该算法是一种局部优化算法。GAVQ算法将遗传算法的全局优化特性和VQ建模技术巧妙地结合起来,采用科学的编码方案,动态的定标技术,高效的交叉策略,是一种全局优化的vQ码本设计算法。并进行了实验检验。

In this paper, a vector quantization speaker model based on Genetic Algorithm is presented. We call it for short as GAVQ speaker model. The algorithm is subsequently called GAVQ algorithm. As vector quantization technic can effectively extract the personal feature, so VQ codebook can be used as a valid speaker model. The traditional method for codebook designing is LBG algorithm, but it is a local optimizing algorithm. Being a global optimizing algorithm, GAVQ algorithm takes scientific encoding method, dynamic scaling technic and high effective crossover strategy. The experiments show that the GAVQ algorithm results in better VQ codebook and performs better time robust. The speaker identification system based on GAVQ speaker model performs higher correct identification rates comparing with that based on VQ speaker model.