汉语字音转换中的多层面多音字读音消歧

Multi-level Polyphone Disambiguation for Mandarin Grapheme-Phoneme Conversion

字音转换问题一直是中文语音合成系统中不可缺少的模块,而多音字消歧是字音转换的核心问题。多音字的词性对于读音消歧有着特殊重要的意义。该文利用词性到读音映射关系将多音字划分为a类和b类。针对不同类别,我们提出一种多层面多音字消歧方案,分别从词性和语义层面上进行消歧,使用决策树模型和手工规则体系对多音字进行处理,实验结果表明,从词性层面上消歧利用决策树模型更好,而手工规则体系在语义层面上消歧更加有效。对每类分别从相应层面进行多音字消歧,正确率从baseline的80.74%达到了96.58%。

In Text-to-Speech （T/S） systems,Grapheme-Phoneme （G2P） conversion is one of the most important modules,and polyphone disambiguation is its key problem.The Part of Speech （POS） of polyphones is very important for pronunciation disambiguation.In this paper,polyphones are divided into two categories according different from mapping from POS of polyphone to pronunciation.A multi-level solution of polyphone disambiguation is proposed for different category of polyphone and the disambiguation is processed on POS level and semantic level separately,on which decision tree model and rule-based system can be applied.The experiment results show that for POS level disambiguation decision tree get better performance while rule-based is more effective for semantic level disambiguation.Polyphones of each class were disambiguated on corresponding level,and the correct rate increased from 80.74% of baseline to 96.58%.