蒙古语声学模型状态聚类:问题集设计

State Clustering of Mongolian Acoustic Models:Design of the Question Set

针对蒙古语语音识别系统急待设计出相应问题集的现状,讨论了蒙古语问题集设计中的原则、音子集的选择、音素表的完善、复合元音的归类、松与紧概念的应用等关键问题的解决方案,给出了一个蒙古语标准音问题集.多次对比实验结果表明,该蒙古语问题集能够有效支持基于决策树的状态聚类与绑定方法.

Aiming at the current situation of lacking question sets in Mongolian speech recogni tion systems, solutions to some key issues in designing Mongolian question sets are discussed, in cluding design principles, selection of phone sets, extension of phoneme tables, classification of diphthongs and applications of slack and tight concepts. Then, a question set suitable for standard Mongolian is presented. Multiple comparative experimental results show that the Mongolian ques tion set is capable of supporting state clustering and tying method effectively based on decision tree.