摘要
为了提高PSPL(position specific posterior lattices)作为语音文档索引时的检索性能,提出一种基于音位属性检测的PSPL改进方法。该方法首先根据信源熵准则找出原始PSPL中不确定度较大的词弧集合,然后利用音位属性对这些词弧集合进行识别结果修正以及后验概率重估,从而实现对PSPL数据结构的改善。实验结果表明,改进后的PSPL在包含更多正确识别结果的同时,解决了后验概率取值不准确的问题,其解码性能和检索性能均优于原始PSPL。
In the research of spoken document indexing, an alternative position specific posterior lat-tices (PSPL) method based on phonological feature detection is proposed to promote the performance of retrieval. Following the information entropy principle, the sets with high uncertainty are firstly found out from the original PSPL, and then the modification of word arc recognition and revaluation of posterior probability are conducted based on phonological features, which improves the PSPL data structures. Experimental results show that more correct recognition results are delivered by the new PSPL, and the problem of incorrect posterior probability is also solved. Furthermore, better perform-ance in terms of decoding and retrieval is obtained compared with the original PSPL.
出处
《信息工程大学学报》
2012年第4期426-431,共6页
Journal of Information Engineering University
基金
国家自然科学基金资助项目(61175017)
关键词
语音文档检索
语音文档索引
PSPL
自动语音识别
音位属性检测
spoken document retrieval
spoken document indexing
position specific posterior lat-tices
automatic speech recognition
phonological feature detection