论文部分内容阅读
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。
In order to improve the performance of automatic pronunciation error detection system in computer aided language learning, a discriminative training method of acoustic model is proposed. This method maximizes the F_1 value of incorrectly detected pronunciations on the correctly labeled non-native speech database as the training parameter of the model parameters. Sigmoid function is used to construct the objective function of F_1 value function smoothly, and the method of constructing weak meaning auxiliary function and parameter updating formula of extending Baum-Welch form are optimized. The strategy of simultaneous optimization of model parameters and phoneme threshold is proposed to ensure the monotonicity of the growth of the objective function. Pronunciation error detection experiments show that this method can effectively increase the training and test data error detection F_1 value. At the same time training data and test data on the accuracy, recall and detection accuracy have significantly improved.