基于区分性特征的音素识别

Phoneme Recognition Based on Distinctive Features

针对当前主流的基于统计模型的语音识别系统没有使用语音产生知识的问题,通过模拟人类的语音感知理解过程提出了一种"自下而上"的基于区分性特征的音素识别方法。该方法首先根据不同音素的发音特点检测得到音素的边界信息;然后利用分类器完成语音的区分性特征检测,并根据区分性特征与音素的对应关系建立映射表;最后利用音素的边界信息得到语音段的特征序列,通过对语音段的特征序列模糊搜索匹配实现音素识别。实验结果表明,相比于传统的基于隐马尔科夫模型的音素识别方法,该方法在识别速度、鲁棒性及可扩展性等方面具有明显优势。

To address the problem that current popular speech recognition systems based on statisti- cal models do not use Speech production knowledge, a ＂bottom-up＂ phone recognition method is proposed based on the distinctive features by simulating the process of human speech recognition. Firstly, the phone boundaries are detected according to the characters of different phonemes; Sec- ondly, the distinctive features are extracted by classifiers, and the mapping table of feature-to-pho- neme is built depending on the distinctive features; Finally, the feature sequences of segments are obtained using phoneme boundaries, and by fuzzy searching and matching through segment features, phoneme recognition is completed. Experimental results show that, compared to the phoneme recog- nition methods based on Hidden Markov Model, this method has prominent advantages in terms of recognition speed, robustness, expansibility etc.