CAJ | 학술논문

由于传统的最大似然训练准则的局限性，区分性训练的研究逐渐成为当前语音识别领域的一个热点，它已经被应用到实际中并且取得了较好的效果。围绕区分性的概念，可以进行多角度的研究，本文的研究的最小音素错误特征训练就是使用最小音素错误训练准则进行特征变换、参数更新以及模型训练。其研究目的在于调整特征，使得目标函数值趋近最优。研究内容包括高维特征向量和变换矩阵，相关实验证明该算法在词正确率方面，相10对于MLE的提升有近3.8%，相对于MPE的提升大概有1.2%。
유우전통적최대사연훈련준칙적국한성，구분성훈련적연구축점성위당전어음식별영역적일개열점，타이경피응용도실제중병차취득료교호적효과。위요구분성적개념，가이진행다각도적연구，본문적연구적최소음소착오특정훈련취시사용최소음소착오훈련준칙진행특정변환、삼수경신이급모형훈련。기연구목적재우조정특정，사득목표함수치추근최우。연구내용포괄고유특정향량화변환구진，상관실험증명해산법재사정학솔방면，상10대우MLE적제승유근3.8%，상대우MPE적제승대개유1.2%。
Due to the limitation of the traditional maximum likelihood training criteria, the research about the discriminative training has become a hot topic in the ifeld of speech recognition. 20 It is applied to practical and has improved the recognition performance. The research is expanded from various aspects. This paper is focus on the feature-MPE. It uses the minimum phone error training criteria to guide the feature transformation, parameter update and model training. The research aims to adjust the feature vector to optimize the objective function. The research includes a high-dimensional feature vectors and transformation matrix. The experiments show the 25 algorithm can improve the correct word rate.(MLE:3.8%. MPE:1.2%).