基于发音特征的发音偏误趋势检测研究

为了提升计算机辅助发音训练(CAPT)系统中发音偏误趋势(PET)的检测效果,确保反馈信息的准确性与有效性,提出一种基于对数似然比的发音特征方法。该方法将多个基于深度神经网络的发音特征提取器用于生成帧级别的对数似然比,然后将对数似然比组成的发音特征用于PET的检测,为学习者提供发音位置和发音方法的正音信息。实验结果表明,发音特征对PET的检测效果优于常用声学特征(MFCC,PLP和f Bank),当发音特征与MFCC特征相结合时,可以进一步提升性能,达到错误接受率为5.0%,错误拒绝率为30.8%,诊断正确率为89.8%的检测效果。

A Study of Articulatory Features Based Detection of Pronunciation Erroneous Tendency

This paper proposed to apply senone log-likelihood ratio based articulatory features (AFs) to improve pronunciation erroneous tendency (PET) detection performance.The feedback information of articulationplacement and articulation-manner could be derived from the definition of PET.The framework of the method involved two main steps.1) A bank of attribute extractors based on neural networks were trained to estimate the log-likelihood ratio (LLR) for each senone at a frame level.2) AFs composed of those LLRs outputted from each attribute extractor were used for detecting PETs.Results demonstrated that the proposed system had better performance than the baseline system using MFCC.Moreover,substantial improvements were obtained by combining AFs with MFCC,achieving a lower false rejection rate of 5.0％,a lower false acceptance rate of 30.8％ and a higher diagnostic accuracy of 89.8％.