论文部分内容阅读
自动发音错误检错中基于最大化F1值的区分性训练方法是最近提出来的一种声学模型训练方法,该方法能够有效增大发音检错系统中的训练和测试数据检错的F1值。对发音质量评估方法上进行研究,提出一种改进的GOP算法来替代传统的GOP算法,改进GOP算法把传统地GOP算法的先求后验概率再求时间归一化改变成先求时间归一化再求后验概率。根据改进GOP算法给出了使用改进GOP算法最大F1准则的参数更新公式,发音检错实验结果表明基于改进的GOP算法的最大F1值准则训练较使用传统的GOP算法具有过训练抑制性好,在训练机上较低的目标函数值上能达到较高的测试集上的F1值等较好的性能。
Discriminant training based on maximizing F1 value in automatic pronunciation error detection is a recently proposed acoustic model training method which can effectively increase the F1 value of training and test data error detection in the pronunciation error detection system. This paper studies the method of pronunciation quality assessment and proposes an improved GOP algorithm instead of the traditional GOP algorithm. The improved GOP algorithm changes the prior probability of the traditional GOP algorithm to seek the time and normalizes the time to first time Then ask for the probability afterwards. According to the improved GOP algorithm, the parameter update formula using the maximum F1 criterion of the improved GOP algorithm is given. The experimental results of the pronunciation error detection show that the maximum F1 value training based on the improved GOP algorithm has better performance than the traditional GOP algorithm. The lower objective function value on the training machine achieves better performance such as F1 on the higher test set.