基于互补子词单元词图融合的集外词识别

Out-of-Vocabulary Word Recognition Based on Lattice Combination of Complement Sub-lexical Units

混合模型在进行集外词识别时,采用不同类型的子词单元通常具有性能上的互补性.基于此种情况,文中提出互补子词单元词图融合的集外词识别方法.首先分别采用音节和字母音素对搭建2套具有性能差异性的混合模型系统.然后获得这2套系统的识别词图,并合并处理词图中的子词单元.最后分别采用基于词图并集和基于词图交集的融合策略融合处理后的词图,得到更好的集外词识别结果.实验表明文中方法性能优于单系统及ROVER方法.

Different sub-lexical units used in hybrid model often provide complementary information for each other during out-of-vocabulary （OOV） words recognition. In this paper, a lattice combination method of complement sub-lexical units for out-of-vocabulary words recognition is proposed. Firstly, two hybrid model systems with performance difference are built respectively by using syllables and graphones. Next, the recognition lattices are obtained from the built systems and the sub-lexical units are preprocessed for the purpose of combination. Finally, the combination strategies based on lattices union and lattices intersection are respectively explored to combine the lattices to acquire the better result of OOV Words recognition . The experimental results show the proposed method is superior to individual system and the recognizer output voting error reduction （ROVER） system in OOV words recognition.