【摘要】 语音识别中常用的HMM/GMM框架由于训练准则和算法的限制,对模式的辨识能力较差;另一种HMM/ANN框架虽具有极强的模式分类能力,但缺乏成熟有效的优化手段。将一种综合两者优点的TANDEM方法应用到普通话发音检错系统中,通过使用区分性训练的神经网络去估计音素级后验概率,经过一系列后续处理将原始MFCC特征转化为TANDEM特征,作为基于HMM统计模型的发音检错系统的输入,进而完成评测过程。实验结果证明,TANDEM方法使系统的检错性能有了较大的提升,结合MLLR等自适应方法的使用效果会更为明显。更多还原

【Abstract】 The HMM/GMM framework which is often used in voice recognition is comparatively poor at mode recognition due to its limitation on training rule and algorithm.Another framework,the HMM/ANN,is better at mode classification but lacks mature and effective optimisation approaches.In the thesis,the authors apply TANDEM which holds the advantages of both frameworks to the mandarin speech error detection system.At the beginning a discriminatingly training neural network estimates rhythm level pro-check probability,then with a series of processes converts the original MFCC feature to TANDEM feature as the input to the HMM statistical model based speech error detection system,consequently completes the evaluation process.Experiment shows the TANDEM approach can greatly improve the error detection performance of the system.It works even better when cooperates with an adaptive approach such as MLLR.更多还原