论文部分内容阅读
为提高自动发音错误检测性能,提出一种区分性特征补偿训练算法.该方法将高斯后验概率矢量经过线性变换后作为偏移量补偿至传统的谱特征.将经过正确度标注的语音数据库上的发音错误检测F1值的最大化作为变换参数的训练准则.推导了目标函数对变换参数的偏导数公式,并利用无约束参数优化例程L-BFGS更新变换参数.发音错误检测实验表明该方法能够有效增大训练和测试集的F1值.并且训练和测试集的精确度、召回率也都有明显提高.在特征优化的基础上进行模型参数训练,检错性能较单独的区分性特征训练、单独的区分性模型训练都有进一步改进.
In order to improve the performance of automatic pronunciation error detection, a discriminative feature compensation training algorithm is proposed in this paper. The Gaussian a posteriori probability vector is linearly transformed and compensated as the offset to the traditional spectral features. The correct labeled speech database The maximum value of the error detection F1 is used as the training criterion of the transformation parameter.The partial derivative formula of the objective function to the transformation parameter is deduced and the transformation parameter is updated by using the unconstrained parameter optimization routine L-BFGS.The pronunciation error detection experiment shows that this method Can effectively increase the F1 value of the training and test set, and the accuracy and recall of the training and test set have also been significantly improved.Training the parameters of the model based on the feature optimization, the error detection performance than the separate distinguishing feature training , Separate discriminative model training has been further improved.