均衡化的改进K均值聚类法

为了进行连续马尔可夫模型的初值提取,提出一种各类在训练样本空间近似均衡分布的K均值聚类法。在聚类的过程中引入惩罚因子,从而限制过多的训练矢量集中于一个或几个类,使样本空间划分近似均匀。连续马尔可夫模型初值提取实验证明,该方法与标准的K均值聚类法、LBG(L inde Buzo G ray)聚类法相比,降低了矢量量化产生的全局失真,各个类在样本空间的分布更加均匀,提高了矢量量化的性能。将该方法用于孤立词识别连续马尔可夫模型的初值提取,可使各个高斯概率密度函数的参数估计更逼近其无偏估计,从而提高了马尔可夫模型初值的可靠性。

Equilibrium Modified K-Means Clustering Method

In order to get the initial values of continuous hidden Markov models,a K-means clustering method that makes the clusters equably distributed in the space of training vectors is proposed.A punishment variable is introduced in the clustering,to limit too many vectors to congregate in one or several clusters.This makes the partition of the samples space more equably.Continuous hidden Markov model initial value experiments proved that the method reduced distortion distance,made the clusters distribute more equably,and improved the performance of vector quantization compared with the standard K-means clustering method and LBG(Linde Buzo Gray) clustering method.This method improved soundness of continuous hidden Markov model initial data in isolated-word recognition.It makes estimates of the parameters of Gauss probability functions approach the agonic estimates.