【摘要】 针对计算机英语语音客观评测给出与专家主观评测相关度的更高结果,提出基于区分性训练的声学模型用以改进客观评测置信分数。首先介绍强制匹配算法得到语音矢量的发音质量评测分数(Goodness ofPronunciation,GOP)的过程,再利用假设检验的数学理论证明基于区分性算法"最小音素错误"训练得到的声学模型比基于传统最大似然算法的声学模型更能得到接近于主观评测的置信分数。通过计算主客观评测结果的相关系数,实验验证了利用区分性声学模型的语音评测系统可以给出更高的置信分数。更多还原

【Abstract】 The primary pursuit of objectively scoring of English pronunciation based on computer-aided evaluation system is more correlated with subjectively scoring made by linguistic experts.The application of discriminative acoustic model on our assessing system to improve objectively scoring confidence is proposed.Firstly,the computational method of correlation evaluation is given.Based on the theory of hypothesis testing,it is proved that the discriminative training algorithm named "Minimum Phone Error"(MPE) can indeed improve the confidence of objectively scoring.Then experiment shows that the discriminatively trained acoustic model can be instrumental in objectively scoring with a higher correlation rate than conventional maximum likelihood(ML) algorithm based one.更多还原