摘要： 针对语音关键词检索中的集外词问题，提出基于最大互信息-最小描述长度(MMI-MDL)的子词集构建算法。根据子词对的互信息挑选聚合对，通过MDL准则判断是否聚合成新的子词。使用该子词集把单词映射成子词的组合用于检索。实验结果表明，与已有的MDL子词集构建算法相比，由MMI-MDL方法得到的子词集对检索性能有较大提高，在相同精确率指标下，集外词的召回率相对MDL算法提高12.1%。

关键词: 集外词, 语音检索, 子词, 最小描述长度, 最大互信息, 词格网络

Abstract: In order to solve the Out-of-Vocabulary(OOV) problem in speech retrieval tasks, this paper presents a construction algorithm of sub-word units based on Maximum Mutual Information and Minimum Description Length(MMI-MDL). It selects candidate pairs according to the mutual information of sub-word pairs, judges whether combining the pairs to a new sub-word through MDL. After getting the sub-word set, map the word into sub-word for retrieval. Experimental results show that compared with the MDL algorithm, the proposed method has a better performance, and achieves a 12.1% relative improvement on the OOV recall rate.

Key words: Out-of-Vocabulary(OOV), speech retrieval, sub-word, Minimum Description Length(MDL), Max Mutual Information(MMI), word lattice network

中图分类号: 

• TP301.6