自动发音错误检测中基于F_1值最大化的声学模型训练方法

Maximum F_1-score acoustic model training for automatic mispronunciation detection

为了提高计算机辅助语言学习中自动发音错误检测系统的性能,提出一种声学模型的区分性训练方法。该方法将经过正确度标注的非母语语音数据库上的发音错误检测的F_1值的最大化作为模型参数的训练准则。采用Sigmoid函数对F_1值函数进行平滑构造目标函数,并利用构造弱意义辅助函数的方法以及扩展Baum-Welch形式的参数更新公式进行优化。提出在模型参数更新与音素门限同时优化的策略保证目标函数增长的单调性。发音错误检测实验表明该方法能够有效地增大训练和测试数据检错的F_1值。同时训练数据和测试数据上的精确度、召回率以及检测正确度都有明显改进。

To improve the performance of automatic mispronunciation detection in computer-assisted language learn- ing, a discriminative acoustic model training method is proposed. The method aims at maximizing the Fl-score of mispronunciation detection results on the annotated non-native speech database. The training objective function is formulated as a smooth form of the Fl-score by using the sigmoid function, and is optimized by using the extended laum-Welch form like updating equations based on the weak-sense auxiliary function method. Simultaneous updating strategy of acoustic models and phone threshold parameters is proposed to ensure monotonicity of the objective function improvement. Mispronunciation detection experiments show that the method is effective in increasing the Fl-score, precision, recall and detection accuracy on both the training and evaluation data set.