基于语音学分类的汉语三音子识别单元的算法  

Phonetic classification-based triphone for continuous mandarin speech recognition

作　　者：李春[1] 王作英[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2003年第1期16-19,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家"八六三"高技术项目(863-306-ZD03-01-2)

摘　　要：为提高语音识别系统的性能,针对汉语语音的单音节结构的特点,提出了建立三音子识别单元的方法。这种方法完全利用语音学知识对上下文进行分类从而实现参数共享,而不同于传统的数据驱动的聚类共享。提出并实现了采用三音子单元的识别系统的训练算法和识别搜索算法。实验表明:基于语音学分类的三音子单元对识别性能有明显的改善,系统的首选误识率相对基线系统降低了28%。This paper proposes a new technique to construct triphones for Mandarin to improve the performance of automatic speech recognition systems. The technique is based on the monosyllabic characteristic structure of Mandarin. As opposed to traditional parametersharing techniques, which are based on datadriven clustering, this method is purely based on phonetics. Phonetics is applied to classify various contexts between syllables to realize parametersharing between different triphone units. A training algorithm is presented with a search network for recognition. Experimental results show that phonetic classification based on the triphone can greatly improve system performance. The proposed method reduces the error rate by 28% compared with a baseline system.

关 键 词：识别单元 汉语连续语音识别 三音子 语音学分类 训练算法 识别算法 音节结构 

分 类 号：TN912.34[电子电信—通信与信息系统]