论文部分内容阅读
隐马尔柯夫模型用最大似然准则训练的结果，能保证训练过程的最佳，但却不能保证识别过程的最佳，从识别时的最小误识率出发导出的各种准则之下的训练方法，能有效地提高系统的性能。本文将校正训练（CT）算法应用于半连续隐马尔柯夫模型（SCHMM）的训练过程，给出了算法的具体实现步骤，同时对于所需的易混集的建立方法，采用一种适于中小词表系统的动态构造方法来实现。
Hidden Markov model training with the maximum likelihood criterion results, to ensure the best training process, but can not guarantee the best recognition process, from the identification of the minimum error rate derived from the training under various criteria Method, can effectively improve the system performance. In this paper, the calibration training (CT) algorithm is applied to the training process of semi-continuous hidden Markov model (SCHMM), and the concrete steps of the algorithm are given. At the same time, The dynamic construction of small and medium vocabulary system to achieve.