【作者】 严可； 魏思； 戴礼荣； 刘庆峰；

【Author】 YAN Ke~1,WEI Si~2,DAI Lirong~1,LIU Qingfeng~(1,2) (1.iFlytek Speech Lab,University of Science and Technology of China, Hefei 230027,China; 2.Anhui USTC iFLYTEK Company Ltd,Hefei 230027,China)

【机构】 中国科学技术大学电子信息工程； 安徽科大讯飞股份有限公司；

【摘要】 帧规整对教后验概率是目前公认最有效的发音质量的度量。但该文证明了该度量还反映了当前发音对应的声学模型与概率空间中声学模型的混淆程度,使其在评分任务上存在着严重的缺陷。因此,该文提出了可训练的音素相关的后验概率变换;并研究了线性和非线性sigmoid变换,同时给出线性变换的显式全局最优解和非线性变换的梯度下降解法。在全国采集的普通话水平测试现场考试的篇章朗读题型上进行实验。实验结果表明:当概率空间包含所有音素时,该方法能使人机相关度从0.582提升至0.768;当采用优化的概率空间时,该方法能使人机相关度从0.696提升至0.773。更多还原

【Abstract】 The frame-normalized log posterior probability is a promising feature for pronunciation evaluation.However,this paper points out its deficiency and proves that this reflects the confusion between the acoustic model of current pronunciation and acoustic models in the probability space.A phoneme-based log posterior probability transformation method is given to deal with this problem using both linear and non-linear transformations with a closed form solution for linear transformations and a gradient descent method for nonlinear sigmoid transformations.Tests on a live Putonghua database indicate its effectiveness.The cross correlation between the human and machine scores increases from 0.582 to 0.768 with posteriors calculated in the all-phoneme probability space and from 0.696 to 0.773 with posteriors calculated in a typical error pattern refined probability space.更多还原