摘要
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。
When trained under the criterion of MLE, the result HMMs could get the best performance for the training data instead of the test data. To overcome this drawback,the criterion which takes account of error rate will benefit the overall performance. In this paper, we propose a method of applying correct training (CT) to semi --continuous hidden Markov model (SCHMM). We also propose a method for construction of confusable word sets.
出处
《航空计算技术》
1998年第4期9-11,共3页
Aeronautical Computing Technique