改进的K均值分割算法在关键词检测中的应用

Application of A Modified Segmental K-Means Training Algorithm to Keyword Spotting

鉴于K均值分割算法中隐马尔可夫模型 (HMM)参数重估公式简单、实用 ,目前大多数基于HMM的关键词检测系统都采用此算法训练参考模型。为了提高参考模型的有效性和解决该算法在具体实现时所遇到的问题 ,本文提出了改进的K均值分割 (MSKM)算法。MSKM算法以关键词检测系统的检出率为模板收敛的判决依据 ,使HMM参数调整从一定程度上而言是以检测系统性能为目标函数 ;同时引入了基于HMM的聚类方法 ,使聚类和参数估计融为一体。实验结果表明 ,采用MSKM算法比原算法可使关键词检测系统的平均检出率提高 1 8%。

As the reestimation formulas for Hidden Markov Model (HMM) parameters in the Segmental K Means training algorithm are very simple and practical,it is widely used to estimate reference models in the HMM based keyword spotting system.In order to improve the efficiency of reference models and solve the problem in implementation of this algorithm,the paper provides a Modified Segmental K Means training algorithm,in which the test for model convergence is made on the basis of the current detection rate of the keyword spotting system,which,to some degree,means the objective function of parameters estimation is directly related to the performance of the system.Meanwhile,the HMM clustering technique is integrated directly into this HMM parameter estimation procedure.The experimental result shows an average detection rate improvement of 1.8% can be obtained by means of the proposed training algorithm.