摘要： 为提高索引覆盖率并获得更多的候选路径，提出一种在词格上融合音位属性的语音文档索引方法。通过基于音位属性检测的语音识别系统建立词格，利用其信息互补性，与传统的词格进行起止节点合并。针对合并后Lattice规模增大的问题，采用基于位置的分段对齐方法对其结构进行压缩。实验结果表明，该方法在提高索引覆盖率和降低最小错误率方面均优于传统的语音文档索引方法，能够有效提高语音检索性能。

关键词: 语音文档检索, 语音文档索引, 自动语音识别, 音位属性检测, 词格, 信息融合

Abstract: A spoken document indexing method integrating phonological feature is proposed to improve indexing coverage and attain more candidate paths. Due to the complementarity between the lattices from different recognition systems, a lattice is firstly generated by the speech recognition system based on phonological feature detection, which then makes nodes merge with the conventional lattice indexing. Furthermore, the segment alignment based on position is employed to compress the scale of combined lattice. Experimental results show that the proposed indexing method can make speech retrieval improvement and have good performance in terms of indexing coverage enhancement and oracle error reduction.

Key words: Spoken Document Retrieval(SDR), spoken document indexing, Automatic Speech Recognition(ASR), phonological feature detection, lattice, information integration

中图分类号: 

• TP391