CAJ | 학술논문

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。
자동발음착오검착중기우최대화F1치적구분성훈련방법시최근제출래적일충성학모형훈련방법，해방법능구유효증대발음검착계통중적훈련화측시수거검착적F1치。대발음질량평고방법상진행연구，제출일충개진적GOP산법래체대전통적GOP산법，개진GOP산법파전통지GOP산법적선구후험개솔재구시간귀일화개변성선구시간귀일화재구후험개솔。근거개진GOP산법급출료사용개진GOP산법최대F1준칙적삼수경신공식，발음검착실험결과표명기우개진적GOP산법적최대F1치준칙훈련교사용전통적GOP산법구유과훈련억제성호，재훈련궤상교저적목표함수치상능체도교고적측시집상적F1치등교호적성능。
Maximum F1-Score Criterion ( MFC) is a newly proposed discriminative training for automatic mispronunciation detection. In the previous work, discriminative training of Hidden Markov Model ( HMM) parameters could improve mispronunciation detection performance. In this paper, a modified form of the GOP calculation which could be more effective in MFC training is presented. The modification is to replace the traditional time-normalized log posterior probability with log posterior of time normalized proba-bility. The MFC model training algorithms and comparison with those using traditional GOP method are de-scribed. Mispronunciation detection experiments indicate that the MFC training with modified GOP is better than that MFC training with traditional GOP.