基于优化检测网络和MLP特征改进发音错误检测的方法

Mispronunciation detection with an optimized detection network and multi-layer perception based features

该文基于优化的检测网络和多层感知(multi-layerperception,MLP)特征,提出一种可以更加准确地检测出错误发音类型的方法。首先,从第二语言学习的语音库中提取出基本的发音规则以及组合的发音规则,并相应地计算它们发生的先验概率,再将这些具有先验概率的规则用于构建基于多发音的扩展检测网络。然后在检测过程中,引入基于发音特征的MLP特征来描述发音概率,替代了传统的语音声学特征。最后使用基于MLP特征的GMM-HMM框架从检测网络中识别出最可能的发音音素串。实验表明:该方法将音素识别正确率提高了3.11%,错误类型准确率提高了7.42%。

This paper describes an optimized detection network for multi-layer pereeptron （MLP） features to more accurately capture mispronunciations. First, the basic and combined phonological rules are extracted from the L2 speech corpus with computation of their prior probability of occurrence. The prior probability rules are then used to build a multiple pronunciation based extended detection network. Then, articulatory based MLP features are introduced to describe the pronunciation probability instead of the conventional speech acoustic features during detection. Finally, the GMM-HMM framework with MLP features is used to pick the most probable pronunciation phoneme sequences from the detection network. Tests show that this approach improves phoneme recognition accuracy by 3.11% and the mispronunciation type accuracy by 7.42%.