在检测出音位属性的基础上,提出了一种基于音位属性后验概率的音素边界检测算法,并将音位属性与边界信息应用于基于条件随机场的音素识别。该方法首先计算得出相邻帧音位属性后验概率向量间的夹、角,然后将夹角的极大值点所在的帧选为候选边界,最后通过约束条件去除极值点中的错误边界。本文将音素边界与音位属性信息进行组合,作为基于条件随机场模型的识别系统的观测特征,实验结果表明,增加边界信息后,音素正确识别率有了显著提升。
A phone boundary detection method is proposed based on the phonological attributes posterior probability, taking these features and boundary information to analyze conditional-random-field-based phone recognition system. Firstly, the angles between posterior probability vectors of adjacent frames are calculated, and then the frames with the maximum angle are marked as the boundary candidates. Secondly, the false phone boundaries are removed through several restrictions in the boundary candidates. Finally, the combination of phonological attrib-utes and phone boundaries is presented as the observation vectors of conditional random fields. Experimental results show that the accuracy rate of phoneme recognition is superior to the base system which only uses phonological attribute features.